Incident automation is mostly an authority problem. A model can suggest a plausible fix; the harder question is whether it should be allowed to act, on what target, for how long, and with what evidence left behind.

This project follows one deliberately narrow remediation path from typed intake through assessment, an immutable proposal, human approval, atomic execution, verification, and audit. Approval grants short-lived authority to perform the exact proposed action. The model never receives open-ended access to the runtime.

What the evidence says

The current evidence includes 114 offline tests, 13 disposable-container tests, and two live integrations. In the strongest end-to-end run, the system took an owned service from HTTP 503 and an unhealthy container state to HTTP 200 and healthy, while preserving the container identity and recording each step.

It is not a production incident-response platform. One bearer token controls mutation, there is no RBAC or production target discovery, and remediation is restricted to disposable resources created by the proof of concept. The container restart is a vehicle for testing approval, telemetry, and authority boundaries, not the product claim.

Read the evidence

• Verification record
• Disposable-service recovery decision
• Security boundary and remaining risks

Once several coding agents are working in parallel, “what should run next?” stops being a prompt and becomes a control-plane question. Chat history is a poor source of truth for dependencies, decisions, stale handoffs, and work that is technically ready but unsafe to dispatch.

Agent Cockpit stores that state in the repository that owns the work. It reads versioned work orders, dependencies, decisions, handoffs, and health checks, then separates hard blockers from degraded signals. A missing approval can stop dispatch; a stale but noncritical signal can warn without freezing everything.

The CLI and synthetic demo exercise the model, including checks against the repository’s own GitHub Actions state. It does not run agents, coordinate writes across repositories, or claim production adoption. Its job is narrower: make dispatch readiness inspectable and reviewable before another agent starts changing code.

Read the design

• Quick start and live-check behavior
• Trust-state dispatch semantics
• Repository-local execution boundary
• Subagent dispatch standard

Valid JSON is only the first boundary. The shape can be correct while the requested tool is unauthorized, its arguments are unsafe, or the tool returns data the rest of the application should not trust.

This runtime puts schemas on both sides of the model/tool boundary. It uses Pydantic validation, a hard attempt budget, typed failures, and an explicit registry that validates tool arguments and results around application-owned code. Plain prose never becomes a tool call, and provider failures stop rather than entering a blind retry loop.

The evidence is intentionally modest. Hermetic tests with scripted models and deterministic tools cover malformed output, recovery budgets, authorization, and result validation. There is not yet a frozen provider-response corpus or enough provider-specific integration evidence to claim broad runtime reliability.

Read the design

• Architecture and attempt semantics
• Test coverage and evidence limits
• Explicit tool-authority decision
• Experiment status

ReAct loops are easy to demo and surprisingly easy to leave unbounded. A model can repeat a plausible action, burn retries on malformed output, or keep planning after it has stopped making progress.

This project wraps observe → decide → act → reflect in a typed state machine. Iterations, tool calls, and structured-output recovery have separate budgets. Exact repeated actions and repeated application-visible state keys stop deterministic cycles. Model, validation, tool, and budget failures return typed summaries rather than escaping as uncontrolled exceptions.

The offline vertical slice covers single-tool and multistep flows. One live local-server run discovered a model and completed an addition tool call. That is useful integration evidence, but it is one narrow path. It does not establish open-ended planning reliability or broad compatibility across providers.

Read the evidence

• Architecture and dependency boundary
• Offline and live experiment record
• Testing boundaries
• Cycle-detection decision

I never got a clean answer. I learned something more useful instead: I couldn’t govern even this small system until I could independently verify its behavior.

Several apparent model failures were something else: a token ceiling, a serving default, an uneven comparison, or a bug in my control flow. The score told me that a model had failed. It rarely told me which part of the system had failed or why.

That became the more interesting result.

The score was right, but the explanation was wrong

Early runs showed a model returning nothing usable. My harness recorded a capability failure.

The raw response told a different story. The model had spent its completion budget on hidden reasoning tokens, hit the ceiling, and stopped with finish_reason: length before producing an answer.

After I raised the limit, 8 of 27 successful calls used more than the old 1,200-token ceiling, almost entirely on hidden reasoning. The model had not necessarily failed at the task. I had cut it off before it could answer.

The pass/fail result alone could not reveal that. I needed two separate things: verification to tell me whether a valid answer existed, and telemetry to explain why it did not.

The model wasn’t the whole system

My most convincing “the stronger model is better” result compared a frontier model running inside a mature coding agent with an open model accessed through a raw API.

The token accounting made the problem obvious:

• Frontier side: 67,991 input tokens
• Open-model side: 1,320 input tokens

One side had an agentic loop, tools, retries, internal turns, and a tuned system prompt. The other received one completion request.

The result was real in the narrow sense that one system completed the task and the other did not. But I had not isolated model capability. I had compared two systems with very different amounts of machinery around them.

Most of that machinery was missing from the result table.

Any useful account of an agent’s behavior has to include the scaffold: which tools it used, how many attempts it received, what information it saw, what the framework changed between turns, and how the final answer was extracted. Looking only at the model leaves much of the actual system unobserved.

My enforcement code was wrong too

The orchestrator was supposed to inspect failing answers and repair them. When I reviewed the control flow, I found that it was rewriting answers before checking whether they were already correct.

The loop was quietly degrading passing results.

Left unfixed, the experiment would have produced a plausible conclusion: orchestration does not help and may make answers worse. The conclusion would also have been wrong. I would have been measuring an ordering bug in the code responsible for running the experiment.

That changed how I thought about the harness. The code enforcing the rules could not be above scrutiny. It needed its own invariants and tests.

Verification had to happen before revision. A passing answer had to remain untouched unless the system could produce evidence that it no longer passed. Every transition needed to be visible.

The checker was part of the attack surface

My checks also admitted shortcuts. A candidate could satisfy the letter of a test by hardcoding a value, exiting early, or bypassing the behavior I intended to measure.

A passing check was therefore not enough. I needed adversarial probes for the obvious shortcuts and a way to mark ambiguous results as invalid rather than quietly counting them as successes.

I’m careful with the word “gaming” here. A model does not need to understand that it is cheating. If the shortest path to a reward exploits a weakness in the evaluator, the effect is the same: the score says one thing while the execution says another.

The verifier is part of the system under test.

What made a result believable

I eventually settled on a few structural rules.

Authority lives in code. The model cannot grant itself more attempts, a larger budget, or additional tools through its output. Those limits hold whether the model understands them or not.

Where an answer can be checked through execution, I run the check in an isolated environment. An LLM judge may be useful for qualitative work, but it should not replace a deterministic test that already exists. Model judges can share blind spots with the outputs they grade.

Ambiguous outcomes fail closed. Hardcoded answers, premature exits, incomplete traces, and unverifiable success claims are not passes.

The orchestration receives the same scrutiny as the model. I record enough about control flow, tool use, budgets, and endpoint behavior to identify which layer produced a failure without logging raw prompts by default.

Runs retain their lineage. Inputs, configuration, decisions, and derived artifacts have enough parentage to reconstruct a result instead of accepting it on faith.

None of these controls depend on the model or its prompt cooperating.

From an eval problem to a control problem

These experiments do not prove a general loss-of-control thesis. They made the underlying engineering problem concrete.

In an eval, weak enforcement or bad telemetry produces a misleading number. When an agent can call APIs, modify a repository, spend money, or act under standing credentials, the same class of failure produces an external side effect.

The questions become more consequential, but they are familiar:

• What was the system allowed to do?
• Which component granted that authority?
• What did it actually do?
• Can the claimed result be checked against external state?
• Does the record remain trustworthy if the model, prompt, or orchestration is buggy, compromised, or simply opaque?

A stronger system prompt cannot answer those questions. The boundary has to exist outside the model: scoped credentials, budgets enforced in code, action receipts, immutable audit records, and independent verification of effects.

Better models still matter. Better orchestration matters too. Neither removes the need for controls that keep working when the system gives you a convincing but incorrect account of itself.

I started out trying to build a loop that made a model perform better. I came away more interested in the layer around that loop: one that bounds its authority, records what it changes, and checks its claims against the world.

I learned that lesson on functions that fit on a screen. The agents won’t stay that small.