# Getting an agent coordinator to brief like a product manager

Author: Chris Chabot
Canonical URL: https://chabot.dev/blog/getting-an-agent-coordinator-to-brief-like-a-product-manager/
Published: 2026-10-18

Why prompt fixes do not hold an agent coordinator's voice across hundreds of turns, and how to split rules between prompt, tool descriptions, state and services so the prompt stays short.

Check in after a few hours away from dozens of coding agents running around the clock, and the update waiting for you too often opens with queue states, test counts, and notes on what the system did not do. Surfacing worker logs, status lines, or a pull request per task is the obvious interface for multi-agent work, and it turns the person steering the project into a log reader, delivering information at the wrong level, volume, and time. A coordinator has to communicate like a product manager briefing a stakeholder, not a coding agent reporting from the weeds: outcomes, consequential choices, and what happens next, leaving implementation detail with the workers.

Between September and November 2026, I designed the communication layer for the project feature of our coding-agent platform, used from a browser or phone. Each project has one long-running coordinator conversation that talks with the project's owner and directs full agent threads, each equivalent to opening a new Claude Code or Codex thread and handing it a large job. The coordinator commissions work rather than executing it, tracking code written, reviewed, landed, and verified in use as separate facts. Even with explicit instructions to stay high-level, 98 of 189 prose replies in our own coordinator thread opened with queue, review, or "did not" wording.

A prompt can steer an LLM's voice for a few turns, but across hundreds of turns saturated with worker reports, test runs, and state dumps, the model regresses to the vocabulary in its context window. Every prompt fix we added held for only a few days before telemetry crept back into the updates. Owner attention is the scarce resource in multi-agent orchestration, so cross-turn rules that protect it belong in typed records and services, not prompt prose. Moving enforcement into services is also what lets the coordinator prompt stay short enough for a model to follow.

## Enforcing the voice in services

Hard constraints on owner-facing prose have to sit at the service boundary. Status updates pass through a typed service that enforces a 120-word cap with a 35-word headline and rejects any draft containing internal IDs, file paths, hashes, test counts, or platform terms such as queue, retry, or receipt.

Relying on the coordinator alone to speak also fails in the exact situations where silence hurts most, such as a stalled turn or a broken worker loop. A service-owned pulse monitors project state independently of the coordinator conversation. It gives the coordinator a grace period to write first, posts an incident update if an issue goes unaddressed for 15 minutes, and guarantees that an active project never goes silent past its digest interval.

Reporting cadence has to follow whether the owner is looking. Presence tracking drives the digest schedule directly to deliver updates that are frequent and detailed while they are watching, only the high-level and critical ones while the project runs in the background, and a prose summary of what happened when they come back. Each digest lists at most five items per category and counts the remainder, a ceiling added after an uncapped digest delivered 785 items in one message. Any coordinator reply consisting solely of internal coordination work with no material change posts nothing.

Escalation operates under the same attention budget; routing every routine trade-off to the owner creates a new bottleneck. Decisions are stored as typed records with explicit options, consequences, and a recommendation. Reversible decisions carry a deadline after which the system applies the recommendation unless the owner intervenes, while standing project policy resolves standard choices automatically, settling about half of 322 decisions in a week before they reached an owner.

## Splitting rules across four layers

It is easy to see why teams default to prompt patching: locally, it works. When a coordinator opens with queue jargon, asks for scope it already has, or leaves worker capacity idle, adding a sentence to the prompt stops that failure on the next run. Across weeks of incident-driven additions, our coordinator prompt grew from about 11,000 to 27,000 characters, and the per-wake protocol injected whenever a worker event or owner message woke the thread swelled alongside it. Contradictions accumulated across sections, including one block requiring an update every few tool calls and another requiring silence when nothing changed. At that size, conflicting rules resolve by recency in the context window, and the model wastes tokens reconciling instructions that should never have coexisted.

Once services enforced the communication invariants, most of that prose was redundant. I rebuilt the coordinator around the base prompt used by standalone agent threads, adding one 41-line section of about 4,500 characters for its management role. The per-wake protocol dropped from about 53,000 characters to 2,900, leaving the state payload intact while cutting paragraphs that restated tool semantics or duplicated service checks.

That reset assigned every rule to the layer suited to hold it:

- **Wake state** holds current project facts.
- **Tool descriptions** hold invocation mechanics, replacing paragraphs of per-tool rules in the prompt with a brief map of tool groups.
- **Services and typed records** hold cross-turn invariants such as output schemas, cadence, fallback pulses, and decision defaults.
- **The prompt** holds only role and judgment.

With enforcement moved out of prose, the coordinator's entire voice rule came down to a single line: high in context, low in detail.

## Auditing the rendered prompt

Cutting a prompt by an order of magnitude introduces its own regression risk: load-bearing constraints disappear, and reading the coordinator section by itself misses conflicts that only surface once the base thread prompt, wake protocol, and tool schemas are assembled at runtime.

After the reset, I ran an independent reviewer agent over the full rendered prompt for every agent role. It found 18 problems the rewrite had left behind, including a remaining copy of the conflict between periodic updates and silence, stale tool names that no longer matched the live schemas, and a lost rule that the owner uses the platform with no source tree or shell and must never be asked to run a command or read a file.
