High-Value If, Low-Value Foreach
Why agents trade in judgment structures, not models
Why model placement, not model frequency, determines whether agents become real products
This is the first in a series on the engineering of Agent Runtimes. It argues that the 2026 problem for agent builders is not intelligence — it is where intelligence is placed. The full series builds the discipline from high-value ifs to value functions, schemas, workflows, and the if/foreach ledger.
Abstract
You are not selling model access. You are selling reusable judgment.
Most agent products fail not because the model is too weak, but because the model is in the wrong place. They put a frontier model inside the foreach loop — re-reading every page, re-planning every step, re-generating every action — then collapse under cost, latency, and reliability failures the moment they leave the demo environment.
The discipline this essay proposes: place the stochastic model at high-value ifs — branches with high uncertainty, dense semantics, and high stakes — and hand the deterministic foreach to programs, workflows, and tools. Once the model makes a useful judgment, the system must absorb it into schemas, workflows, audit logs, skills, and evaluation sets. The goal is not more model calls. The goal is fewer unnecessary ones over time.
Frontier models open paths. Deterministic systems pave roads. Everything else is engineering.
1. The 2026 Problem for Agent Builders
Most agent demos — built by what I’ll call agent builders: the teams designing model routing, workflows, schemas, permissions, approvals, and audit logs — fail for the same reason: they put the model in the wrong place.
They ask the model to read every page, re-understand every DOM, re-plan every step, and generate every action. The demo looks impressive. The product collapses under cost pressure, latency, reliability failures, and lack of reuse.
The first commercial problem of agents is not intelligence. It is where intelligence is placed. Where should the stochastic model appear inside a deterministic system?
Over the past two years, from the Chat Completions explosion to the emergence of Computer Use compatibility layers, the industry has hit a structural wall: there is no mechanism to crystallize stochastic judgment into deterministic assets. Every model call is still a one-time expense. When the task ends, the exploration vanishes with it.
In 2026, the model layer is moving from chat-completion interfaces toward agent-building primitives — Responses API, Structured Outputs, Computer Use, Agents SDK, Sandbox Agents. Frontier inference costs are dropping into a range where deep semantic analysis finally pencils out. Local runtime ecosystems are becoming mature enough to orchestrate complex tasks. And the old internet spent three decades compressing reality into low-resolution proxies — degrees, ratings, sales volume, clicks, rankings, and other sortable fields — which are now exposing systemic weakness under information degradation and Goodhart’s Law.
Agent builders need a complete methodology spanning high-value ifs, value functions, and engineering implementation. The question reduces to one thing:
How can a single model judgment become a durable system asset?
Builders won’t win by out-scaling model providers. They’ll win by owning the judgment structures, workflows, and responsibility boundaries that models serve.
2. Model Placement, Not Model Frequency
In agent systems, model placement matters more than model frequency.
The do-everything agent that wows demo audiences dies in production from cost blowups and reliability collapse. The answer lies not in parameter count or prompt sophistication, but in the most fundamental engineering discipline: where the model sits in the architecture.
Luo Fuli, who led DeepSeek-V2 research before joining Xiaomi’s MiMo team in late 2025, framed it in a recent interview as a product of three factors: effectiveness × cost × efficiency. The framing matters because it forces agent design back from the “fully autonomous” myth to the engineering reality of the runtime. None of the three is optional. Effectiveness without cost control cannot become a commercial product. Low cost without trustworthy results creates no willingness to pay. Efficiency that only runs deterministic flows is not an agent — it is RPA with a wrapper.
Translating this into concrete architectural decisions gives us the first engineering discipline of the agent era.
What Models Should and Should Not Do
The past two years of Agent Runtime practice — the execution layer that manages state, permissions, workflows, tool calls, model routing, audit logs, and human approval — point repeatedly to one division of labor:
The Agent Runtime should place the model at high-uncertainty, semantically dense, high-stakes decision branches. It should hand repeatable, well-scoped, verifiable, low-risk execution to programs, workflows, and tool layers.
Low-value foreach: what it looks like
Foreach is batchable, repetitive, procedural, well-scoped work:
- Batch-scraping web pages
- Batch-organizing files
- Batch-exporting conversation logs
- Batch-converting Markdown to other formats
- Batch-executing known browser flows
- Batch-calling API endpoints
- Batch-structuring content into uniform formats
- Batch-classifying items in a low-risk first pass
These tasks are often packaged as “AI automation,” but structurally they belong to programs, scripts, CLI, SQL, queue systems, rule engines, Playwright, MCP tools, or deterministic pipelines.
On tasks that can be deterministically described, programs are typically orders of magnitude cheaper, faster, more stable, and easier to audit and reproduce than LLM calls.
If a task can be described as:
for item in items:
do_something(item)
the model should not be the primary executor of that loop.
High-value if: the model’s anchor point
I call these branches high-value ifs: decision points with high uncertainty, semantic density, and high stakes. They cannot be reliably captured by a few fields, nor safely decided by low-trust rules.
Not every if is high-value. file_exists(path) is an if — but it is a state check. user.is_logged_in is an if — but it is a boolean lookup. inventory > 0 is an if — but it is a numeric comparison. These are low-value state judgments better handled by programs, databases, caches, rule engines, or workflows.
High-value ifs in an Agent Runtime are judgment nodes that require the model to synthesize context, objectives, risk, constraints, and user intent across multiple possibilities, then choose:
-
Intent judgment: What does the user actually want?
-
Information sufficiency: Is the current context enough to proceed?
-
Cost-benefit judgment: Is this task worth continuing?
-
Tool selection: Which tool best fits this step?
-
Quality judgment: Is this result trustworthy?
-
Signal-noise judgment: Which parts are noise?
-
Information value: What should enter long-term memory?
-
Risk classification: Which operations need human confirmation?
-
Pattern recognition: Which processes can be hardened into workflows?
-
Error attribution: Is the failure a page change, permission issue, model misunderstanding, or unclear user intent?
None of these can be solved by a simple foreach. These decision branches are where the LLM should appear in an agent system.
A high-value if is not completed by the LLM alone. It is completed by two layers:
Layer 1 — The LLM operates in high-dimensional semantic space: understanding, extracting, attributing, comparing, generating candidate judgments, and handling uncertainty.
Layer 2 — Schema, enums, booleans, workflows, rules, and human approval gates constrain the model’s output into auditable, reproducible, accountable final decisions.
Without the LLM, high-value ifs regress to field filtering. Without the schema, high-value ifs become unauditable free-form generation. Both layers are indispensable.
If and Foreach Are Nested, Not Exclusive
High-value ifs and low-value foreach loops rarely appear in isolation. They interleave constantly.
The user says: “Organize this week’s important emails and classify them by project.” Structurally, this is a foreach:
for email in this_week:
classify(email)
But expand one layer, and every step contains ifs:
for email in this_week:
if is_spam(email):
skip(email)
if is_newsletter(email):
archive(email)
if is_work_related(email):
project = infer_project(email)
if project.confidence > 0.85:
classify_into_project(email, project)
else:
ask_user_or_create_candidate_project(email)
if contains_deadline(email):
date = extract_date(email)
if date.confidence > 0.9:
create_task(email, date)
else:
mark_for_review(email)
If and foreach nest inside each other. This nesting does not imply every step should call an LLM. The primary engineering job of the Agent Runtime is to precisely decompose this mixed structure:
- Which foreach loops can run through deterministic pipelines with no LLM at all.
- Which ifs require a frontier model, because they are high-value judgment nodes.
- Which ifs can be resolved by historical workflows, because the pattern has been absorbed.
- Which ifs can be handled by a lightweight model plus rule engine, because the risk is low.
- Which ifs require human approval, because the action is irreversible.
- Which exceptions need to re-invoke a frontier model, because existing workflows have broken.
The key thesis compresses to one line:
The key to agent engineering is model placement.
3. Why Agents Fail When the Model Becomes the Executor
Many agent products fail because they treat the LLM as a universal executor rather than a critical judgment engine. They make the model re-read the page, re-understand the DOM, re-plan the path, and re-generate the action at every step. This looks powerful in demos. Once deployed, it hits five structural failures simultaneously.
Failure 1: Uncontrollable cost
Every step calls a frontier model and recomputes the full context. A task advertised as “fully automated” rapidly accumulates model cost, latency cost, and failure-recovery cost.
Failure 2: Uncontrollable reliability
An LLM is a stochastic system. Every re-decision adds another stochastic failure point. A 10-step task at 95% per-step success has ~60% overall success. A 20-step task drops to ~36%.
The 95%^n figure is an independence approximation. Real agent steps are not independent: page changes, permission expirations, context pollution, and upstream errors increase correlation, making actual failure modes more complex. The problem is not that the model is “slightly off.” It is the physics of error accumulation in a long-sequence autoregressive decision process.
Failure 3: Uncontrollable state pollution
As the LLM ingests execution results across a long loop, the context gets progressively dirtier. Early system prompts, task objectives, risk boundaries, and output constraints get diluted by page noise, failed attempts, and intermediate results. Lost-in-the-middle (U-shaped attention decay in long contexts, Liu et al., 2023) amplifies state pollution. Handing foreach to programs means every necessary model call happens inside a shorter, cleaner context window.
Failure 4: Throughput-latency mismatch
Foreach means high throughput demand. If means high latency tolerance. Batch-processing thousands of items requires throughput, concurrency, retries, and monitoring. LLMs’ autoregressive generation makes them inherently unsuited to high-throughput execution. A model can be slow and expensive at a few critical nodes, but not at every iteration of a pipeline.
Failure 5: Non-reusable results
Every task starts from scratch. After the task ends, the system retains nothing. The same user issuing the same task triggers the same cost and probability gamble from zero.
These five failures explain why many Computer Use and Browser Agent products stall: they put the LLM inside the foreach, then try to brute-force their way out by reaching for stronger models. The root cause is failing to draw the physical boundary between stochastic judgment and deterministic execution.
The model contributes semantic judgment at the high-value if. The system crystallizes that judgment into structure. The foreach then reuses it at low cost.
4. Explore, Prune, Harden, Evolve
Expanding “model placement over model frequency” into a product architecture yields four phases.
Phase 1: Explore
The user initiates a new type of task for the first time. The system recognizes an unseen scenario: no reusable workflow, no cache hit, no known fallback.
This is the moment to call the frontier model.
Its job goes beyond “completing this task.” It explores the real structure of the task in high-dimensional semantic space:
- What is the user’s true intent — does it match the literal request?
- Where are the critical judgment nodes?
- Which steps are deterministic, which require judgment?
- What are the success conditions for the critical path?
- Where should the system fall back on failure?
- Which steps involve irreversible actions requiring human approval?
The frontier model’s output should be a structured decision graph: explicit schema, explicit step sequence, explicit branching conditions, explicit success signals, explicit fallbacks.
This call is expensive because it bears the cost of task-structure discovery, risk-boundary identification, and the foundation for all subsequent low-cost reuse.
Phase 2: Prune
The decision graph from Phase 1 is a first draft — “theoretically completable.” Real environments expose stability gaps: pages redesign, networks fluctuate, user inputs are ambiguous, third-party APIs fail intermittently.
The key action is testing and pruning with real data:
- Which branch conditions have never actually been triggered.
- Which “requires human review” nodes can actually be automated.
- Which nominally uncertain steps always take the same branch in practice.
- Which fallback paths are more stable than the primary path.
Pruning compresses the “redundant decision graph” into a “minimal necessary decision graph validated by real data.” This phase often does not need a frontier model. Mid-tier models, real-data replays, rule-based statistics, and manual review are usually sufficient.
Phase 3: Harden
The pruned decision graph gets hardened into an executable structured workflow. A real industrial Agent Runtime stores DAG-level (directed acyclic graph) or FSM-level (finite state machine) state-transition logic — “under what conditions to transition from state A to state B,” not “do step one then step two.”
A simplified example:
{
"workflow": "export_bank_statement",
"site": "example_bank.com",
"intent": "download monthly statement as PDF",
"preconditions": [
"user_logged_in",
"two_factor_completed"
],
"steps": [
{
"type": "navigate",
"target": "/account/statements"
},
{
"type": "click",
"selector_hint": "button containing 'Statements'"
},
{
"type": "select",
"field": "month",
"value_from_user": "target_month"
},
{
"type": "click",
"selector_hint": "download_pdf_button"
}
],
"approval_required": false,
"success_signal": "file_downloaded_with_pdf_extension",
"fallback": [
"ask_user_to_confirm_page",
"rerun_page_exploration"
],
"max_retries": 2,
"confidence": 0.78
}
Key properties of this workflow:
-
Structured and machine-parseable — not a natural-language wish.
-
Explicitly declares uncertainty —
selector_hint instead of pretending there is always a stable selector.
-
Explicit success signal — not a false completion.
-
Fallback paths — knows where to retreat on failure.
-
Retry limits — prevents infinite loops.
-
Confidence score — knows how reliable it currently is.
From LLM output to stable DAG, there must be an alignment layer. Real hardening is not saving model-generated JSON directly into production. It means decomposing successful trajectories into nodes, actions, preconditions, state transitions, failure branches, and success signals, then validating through replay, unit tests, shadow testing, and evals. Any action hardened into a workflow must first pass deterministic shadow testing or evaluation-set replay. The confidence score is computed, not self-reported by the model.
The second time, the tenth time, the hundredth time the user says “export my statement,” the system does not re-invoke the frontier model. It runs the workflow deterministically, supplementing judgment only at local uncertainty nodes. Marginal cost approaches zero.
Phase 4: Evolve
Workflows are not one-time artifacts. Pages redesign, DOMs change, CAPTCHAs appear, A/B tests show different interfaces, sessions expire, permissions shift. All of these can break an existing workflow.
The system should not immediately invoke a frontier model for full re-exploration. It should follow an escalation gradient:
- Try existing fallback paths.
- If fallbacks also fail, trigger a limited-scope LLM re-judgment.
- If the failure mode is known, update the corresponding workflow branch directly.
- If the failure mode is novel, escalate to the frontier model for re-exploration.
The key variable across all four phases is confidence. A workflow is a set of judgment assets with confidence scores, not rigid rules. Each successful execution raises confidence. Each failure lowers it. When confidence drops below a threshold, the system re-invokes model exploration.
Combining the four phases:
The model opens the road. The system paves it.
The cost curve shifts from “linear growth per model call” to “growth per new scenario and exception.” The former gets more expensive over time. The latter gets cheaper.
The asset is not the answer
The asset is not that the model answered once. The asset is the absorbed judgment.
If the model’s output cannot be captured by the system, even an excellent answer is just an expensive one-time consumption. If the model’s judgment can be absorbed, even an expensive call is an asset investment — it becomes next time’s rule, workflow, skill, or memory, lowering the cost of the next hundred tasks.
- A correct if becomes a rule.
- A stable branch becomes a workflow.
- A recurring judgment pattern becomes a skill.
- A persistently useful user preference becomes memory.
The dividing line between chat products and Agent Runtimes:
Chat products optimize single-answer quality. Agent Runtimes optimize the conversion of judgment into reusable system assets.
The effectiveness × cost × efficiency formula translates into precise engineering:
-
Effectiveness: determined by the frontier model’s high-dimensional semantic judgment quality during first exploration.
-
Cost: determined by whether the system can absorb that judgment into a workflow.
-
Efficiency: determined by the stability of downstream foreach execution.
These three are multiplicative, not additive. Great effectiveness with uncontrolled cost cannot become commercially viable. Low cost with untrustworthy results creates no willingness to pay. High efficiency that only runs deterministic flows is just RPA.
The engineering discipline:
- Put high-value ifs on the frontier model — protect effectiveness.
- Absorb successful ifs into workflows — protect cost.
- Put foreach on deterministic execution — protect efficiency.
All three are necessary.
5. A Schema Is a Boundary, Not a Form
The four-phase cycle leaves one critical question: what exactly is an if node inside a workflow?
Many builders, the first time they see a JSON workflow or structured schema, instinctively read it as “field filtering from traditional software engineering”:
if degree == "Bachelor's" and years_of_experience >= 5:
pass
else:
reject()
This reading immediately degrades the agent into “SQL queries wearing an AI costume.”
A mature agent if node is a boundary condition of a high-dimensional judgment space actively defined by the builder — not field filtering.
A schema is a directional constraint on LLM compute. An LLM without schema constraints is a high-dimensional hallucination field. A well-designed schema projects a precise decision mold into the LLM’s latent space. The builder designs that mold.
If a content-filtering agent’s output enum contains only LIKE, DISLIKE, SKIP, it structurally cannot express “this article creates cognitive friction, but it is important.”
If a research agent’s output enum contains only USEFUL and USELESS, it structurally cannot express “this material is not directly usable, but it provides a critical counter-example.”
If a recruiting agent’s rejection reasons include EDUCATION_BACKGROUND_TOO_LOW, the agent structurally inherits degree-based elimination. If that enum value is removed, the agent structurally cannot use “degree too low” as a final rejection reason.
The boundary of the enum is the boundary of the agent’s thought.
From field filtering to high-dimensional optimization
In a recruiting scenario, a recruiting agent’s schema should not default to a field filter of “degree, years of experience, tech-stack keywords.” It should be a judgment space with constraints, dynamic weights, and evidence chains.
Context includes: industry profile, real job requirements, reporting line, compensation budget, team structure, and candidate status. Dynamic weights can include: degree, big-company experience, project depth, technical capability, chaos tolerance, and output quality under extreme pressure. Hard constraints include: salary must not exceed budget ceiling, availability must not be later than project start date. Output structure includes: interview recommendation, rejection reason, hidden-potential flag, human-review escalation, and evidence chain.
The critical point is the rejection-reason enum. It may include: core skill gap, excessive job-hopping, budget overrun, availability mismatch, cultural fit risk. But it must not include EDUCATION_BACKGROUND_TOO_LOW.
First, the agent can treat education as a dynamic signal but cannot use it as a final rejection conclusion. Education may carry weight in stable organizations and approach zero in an early-stage startup. It is a variable, not a verdict.
Second, compensation budget is a hard constraint, not a dynamic weight. Hard constraints cannot be compensated by other dimensions. Even if a candidate is outstanding, the agent cannot bypass the budget limit — it must escalate to human review.
Third, the agent performs constrained high-dimensional optimization, not field matching. The LLM compresses the semantic world into optimization parameters — extracting dynamic weights w_i, candidate feature scores Feature_i, evidence chains, constraints, and uncertainty. Ranking, filtering, combinatorial selection, and budget-constraint solving are handled by deterministic algorithms.
The system does not solve:
if degree_ok and years_ok:
interview()
It solves:
Find candidate c that maximizes Σ(w_i * Feature_i)
Subject to:
Salary <= Budget
Availability <= ProjectStartDate
Irreversible risks escalate to human review
Key conclusions require evidence_chain
The dynamic weights w_i are the critical lever. Education should not become an automatic elimination criterion — not because a prompt says ignore_degree = True, but because the schema forces the agent into a larger context for trade-off analysis: company stage, real role needs, team gap, budget rigidity, and whether the candidate’s strengths compensate for weaknesses.
The shift from field filtering to scenario-level judgment depends on structural constraints, not prompt instructions.
For 25 to 30 years, platforms across recruiting, e-commerce, content, healthcare, and investment have relied on low-resolution proxies — degrees, sales volume, ratings, rankings — not because they preferred these fields, but because they lacked scalable high-dimensional judgment capability.
A schema is the builder’s boundary definition of a high-dimensional judgment space. It is not a field table.
6. Every High-Value If Encodes a Value Function
The preceding sections solve “who judges” and “when to judge.” They do not yet solve “on what basis.”
If the builder only says “hand this to the frontier model,” a gap remains: what ordering does the model’s output obey? What matters more? What cannot be breached? When should the system stop and hand responsibility back to a human?
Every high-value if needs an explicit value function behind it. If you do not define the value function explicitly, the model’s default training biases, the platform’s default commercial interests, or random noise will define it for you.
By “value function,” I do not mean a narrow reinforcement-learning reward function. I mean an operational decision structure: objectives, dynamic weights, hard constraints, risk thresholds, rejection rules, human-approval conditions, and evidence requirements.
The if is the entry point of the value function
Every high-value if is a moment when the product makes a trade-off among multiple values: what matters more, what can be sacrificed, what must stop, what is worth recommending, what must be rejected.
In the old software era, value functions were hardcoded by PMs, algorithm engineers, and platform commercial objectives. Recommendation systems decided dwell-time weight, completion-rate weight, click-through-rate weight. Recruiting platforms decided how degree, company background, tenure, and keywords entered ranking. Users entered the system, provided data and clicks; rules were defined by the system.
The power shift in the agent era: from “system defines rules, user provides data” to “users express value-function parameters, preferences, constraints, and risk boundaries in natural language, and the agent compiles them into executable if/foreach structures that the runtime can run.”
- A user says “screen suppliers, but prioritize eliminating delivery risk over low price” — injecting a procurement value function.
- A user says “screen candidates, don’t use education as a rejection reason, focus on real output in chaotic environments” — rewriting the recruiting agent’s judgment boundary.
- A user says “review this contract; the deal can move slowly, but liability transfer and data-usage boundaries need to be conservative” — defining the contract agent’s risk function.
The high-value if is the entry point through which the user’s value function enters the Runtime. The agent’s job is to compile these natural-language value orderings into schema, weights, hard constraints, human-review rules, audit fields, workflow branches, and executable tool calls. Natural language is not the endpoint. The Runtime is the endpoint. Expression is not the asset. A compiled, executed, and reviewable value function is the asset.
The intent compiler is not a shallow metaphor for “translating user speech into code.” What it truly compiles is value functions: pressing vague, implicit, trade-off-laden user intent into structures the system can execute, audit, roll back, and reuse.
No high-value if is neutral
There is no such thing as a neutral high-value if. What appears to be a “neutral tool” often simply hides its default value function.
Internet platforms make daily ifs. Who ranks first in search? What does the feed show next? Which product appears in e-commerce search? Which candidate gets exposure on the recruiting platform? These ifs look like algorithmic ranking. They are actually value ranking. The most common default value function for internet platforms has been maximizing dwell time, click-through rate, conversion rate, ad revenue, and GMV (gross merchandise volume).
Current mainstream LLMs have ifs too. When the user asks a question — give a complete answer, or admit uncertainty? When the user’s plan has problems — agree, or point it out? When the user is emotional — empathize, or encourage calm? These are not tone choices. They are value choices.
Mainstream models have established Provider-level Alignment: general behavioral norms, safety rules, preference alignment. But Provider-level Alignment cannot replace Builder-level, Domain-level, or User-level value functions. The model company defines the lowest-common-denominator behavioral boundary. Industrial agents need concrete value orderings for specific scenarios.
High-value ifs often have a trolley-problem structure: different value frameworks can be internally coherent while producing different decisions. The issue is not that the model is not smart enough. The issue is that the decision itself has no value-neutral answer.
Factual boundaries come before value functions. Factual errors are still errors. Violations of law are still violations. Safety incidents are still incidents. Facts, law, safety, and the responsibility floor cannot be rewritten by value functions. High-value ifs enter the choice space only above these floors.
A stronger model cannot define the value function for the builder
A common approach is a consultant-executor architecture: a cheaper model handles execution, consulting a stronger model at critical points. This reduces cost and improves task completion. But it solves the capability-and-cost problem, not the value-function-source problem.
The stronger model’s judgment still comes from the provider’s blended alignment and default value function. It does not automatically know what your product stands for, what your user’s authorization boundaries are, what your responsibility structure looks like, or which bottom lines you would rather lose short-term conversion to defend.
The builder’s responsibility is not to find a stronger model to make value judgments. It is to actively define: in this specific scenario, what value function governs the high-value ifs — and then ensure that regardless of which model tier is called, this value function is caught by schema, workflow, audit log, and human-approval boundaries.
Engineering solutions can optimize capability. Engineering solutions cannot assume the builder’s value-function definition authority.
An industrial agent must be delegable, auditable, reproducible, and accountable over time. This is why the LLM cannot be the final decision structure. It handles ambiguity; the runtime collapses ambiguity into schema, workflow, approval gates, and audit logs.
The engineering triad: schema, workflow, audit log
The builder cannot write a value function into an agent with slogans. Three instruments are needed:
Schema is the shape of the value function. It specifies which signals are usable, which are reference-only, which constraints are hard, which conclusions are allowed, and which judgments must escalate to human review.
Workflow is the path of the value function. If the product values long-term interest, the workflow must read a long time window. If the product opposes preference flattery, the workflow must surface opposing perspectives. If the product respects factual boundaries, the workflow must perform cross-source verification and explicitly label uncertainty. The LLM cannot skip risk enumeration because the user is in a hurry, omit alternatives because the answer sounds fluent, or pretend information is sufficient when it is not.
Audit log is the evidence of the value function. It records authorization scope, input data, model version, schema version, workflow path, human-review nodes, final action, and post-hoc remediation. An agent without audit logs cannot enter high-trust workflows: errors cannot be located, located errors cannot be explained, unexplained errors cannot be fixed, and unfixed errors cannot sustain long-term responsibility.
Schema gives the value function structure. Workflow makes it reproducible. Audit log makes it accountable.
The Value Function Spec
Every high-value if can be written as a one-page Value Function Spec: primary objective, hard boundaries, what may be sacrificed, what is never sacrificed, who receives escalation on conflict, and which output-schema fields express evidence, gaps, counter-examples, and rejection reasons. Without this spec, “intelligent judgment” is just model default bias.
Taking supplier screening as an example:
-
Metadata: version, domain, responsible party, floor constraints. The system serves buyer procurement, not supplier exposure. Compliance, no fabricated evidence, and no hidden paid ranking are floor constraints.
-
Objectives and dynamic weights: the primary objective is minimizing total procurement risk, not finding the cheapest supplier. Delivery reliability, quality-evidence strength, total cost of ownership, and compliance cleanliness outrank the quote. Low price is a weak signal. Buyer long-term health holds veto power.
-
Prohibited decision factors: supplier paid exposure, platform ranking, and sales-pitch quality cannot serve as recommendation or elimination criteria.
-
Hard constraints and escalation: high compliance risk requires human review or rejection. Missing core evidence requires supplementary materials. Amounts above the approval threshold enter Human-in-the-loop approval. Hard constraints cannot be compensated by other strengths.
-
Output structure: finite decision enum, confidence score, evidence chain, counter-evidence, missing-information type, risk flags, next-step recommendations, and human-review flag. missing_information_type and counter_evidence must be in the output structure.
-
Workflow and audit: the workflow enforces evidence collection, material normalization, dynamic-weight scoring, hard-constraint checking, counter-evidence generation, human-review routing, and audit writing. The audit log records value-function version, input source, model version, workflow version, overridden recommendations, missing-information types, confidence scores, and human-review records.
The engineering chain:
Thought → Schema → Workflow → Audit Log.
Thought that does not enter schema has no structure. Schema that does not enter workflow cannot be reproduced. Workflow that does not enter audit log cannot be held accountable.
A value function is not philosophical decoration. It is the judgment structure of the Agent Runtime.
Restraint as engineering discipline
A high-trust agent is defined less by what it can do than by what it refuses to do: actions it can automate but should not, revenue it can capture but should not, preferences it can flatter but should not, and ambiguity it can hide behind but should not.
Can act, but does not: capability restraint enters the workflow. A recruiting agent can auto-decide whether a candidate enters the final round, but senior positions should escalate to human review. An investment agent can auto-generate buy/sell recommendations, but decisions exceeding a percentage of user assets need human approval. A legal agent can assess contract risk, but major legal liability requires a professional’s signature. The implementation is physically inserting Human-in-the-loop approval nodes into the workflow. When contract amounts exceed the threshold, user asset exposure exceeds risk budget, or an action creates an external commitment, the process pauses and requests human approval — and this node cannot be bypassed by the model, user emotion, or commercial metrics.
Can profit, but does not: metrics restraint enters the schema’s dynamic weights. The product could add more aggressive recommendations, stronger emotional hooks, harder-to-cancel flows — all boosting short-term metrics. But a high-trust agent cannot turn user authorization into growth fuel. The implementation gives user_long_term_health, trust_preservation, and downside_risk weight or veto power that overrides short-term GMV in the schema. Once addiction risk, hidden costs, or irreversible financial loss triggers, the system stops the conversion path.
Can flatter, but does not: emotional restraint enters the workflow branch. The user wants their startup direction affirmed — the agent still points out major risks. The user wants to make an impulsive investment — the agent still opposes. The user is emotionally agitated and wants to decide immediately — the agent triggers a cooling-off period. Implementation means writing emotional-polarization detection into workflow branches. When panic, euphoria, or anger is detected and the current decision is irreversible, the system enters a cooling period, generates counter-arguments, and requires delayed confirmation.
Can obscure, but does not: responsibility restraint enters the audit log. Vague answers are safest: “both sides have a point,” “it depends,” “this is complex, consider everything.” These rarely err and rarely help. An industrial agent should give clear judgment when evidence is sufficient, explicitly acknowledge uncertainty when evidence is lacking, and explicitly escalate to human review when responsibility exceeds its boundary. Implementation means making “I don’t know” leave computational evidence: the system records decision state, missing-information type, confidence score, evidence-gap reason, and next-step action.
Restraint that stays in text gets swallowed by growth pressure. Restraint that enters schema, workflow, and audit log becomes systemic discipline.
7. The If/Foreach Ledger
Model placement ultimately needs to materialize as an if/foreach ledger.
Every agent MVP needs to maintain this ledger. It records at least eight categories:
- Which nodes are high-value ifs, and why they cannot be handled by fields, rules, or low-trust models.
- Which nodes are low-value ifs, and why they can be handed to programs, rules, small models, or caches.
- Which steps are foreach, and whether they can be batched, parallelized, retried, replayed, and monitored.
- Which ifs have already been absorbed by historical workflows and do not need re-exploration each time.
- Which nodes need a frontier model to open the road, and which only need a mid-tier model to assist.
- Which nodes need human approval, because the action is irreversible or responsibility cannot be automatically transferred.
- Which successful judgments can be encoded as workflows, skills, memory, or evaluation samples.
- Which downgrade conditions have been met, allowing expensive model calls to be moved off the critical path.
The ledger’s purpose is not documentation theater. It forces the builder to ask before every model call: is this truly incompressible uncertainty, or has the system lazily wrapped repetitive execution in an “intelligence” label?
How the ledger matures
The first time, the ledger records where the task genuinely requires judgment.
The second time, it records which judgments can be absorbed by workflows.
The third time, it records which processes have migrated from model calls to deterministic execution.
By the tenth time, the ledger starts exposing product boundaries: which judgments still need humans, which can be entrusted to models, and which actions should never be automated.
A mature Agent Runtime is not one that calls the strongest model everywhere. It is one whose ledger gets thicker over time: fewer unnecessary model calls, more hardened workflows, clearer approval gates, and a better understanding of which decisions should never be automated.
Once model placement is ledgered, value functions, low-resolution proxies, and responsibility redistribution all become engineerable problems.
Summary
The agent shift is not only a model shift. It redistributes computation, responsibility, and asset ownership.
The core concepts of this essay:
-
High-value if / low-value foreach: the model appears at judgment branches, not inside execution loops.
-
Model placement over model frequency: the key question is not how many times the model is called, but where it is placed.
-
The model opens the road, the system paves it: explore → prune → harden → evolve.
-
A schema is a boundary, not a form: the boundary of the enum is the boundary of the agent’s thought.
-
Every high-value if encodes a value function: no high-value if is neutral; someone always defines what matters more.
-
Schema / workflow / audit log: thought must enter schema to have structure, enter workflow to be reproducible, enter audit log to be accountable.
-
The if/foreach ledger: the engineering artifact that turns abstract discipline into daily practice.
An agent is, at its core, an intent compiler. The old internet compressed human intent into low-resolution SQL through forms and clicks. In the agent era, users express intent in natural language and ambiguous context. The agent’s entire job is to first identify where incompressible uncertainty exists, then compile the clarified portions into a tightly-fitted, responsibility-bounded foreach automation pipeline. The agent absorbs uncertainty at compile time. The system sprints on deterministic rails at runtime.
In the long run, whoever defines the high-value ifs controls the decision layer. Whoever controls the decision layer builds a practical form of digital sovereignty — not through owning the model, but through owning the judgment structure that the model serves.
The builder’s position comes down to this: in every specific scenario, write a clear, explainable, reproducible, and accountable value function into every high-value if.
