Agent Harness Blueprint

Multi-agent systems that don't fall apart after five messages.

A developer reference for building agent systems that hold together in production. Three-layer architecture, five annotated system prompt templates, a structured harness loop, state management patterns, and a failure mode playbook. $29, instant download.

Get the blueprint — $29
Markdown files · instant download · framework-agnostic

What's inside

Reference architecture Three-layer design (lead, worker, specialist) — what each layer owns, how communication flows, when to add a layer vs. a tool
5 system prompt templates Lead orchestrator, background worker, research agent, QA/verifier, code reviewer — each annotated with the reasoning behind every design choice
Harness loop pattern Plan → execute → review → iterate. How to pass context between stages, when to loop vs. stop, how the verifier stage prevents broken work from shipping
State management What to persist to disk vs. keep in context, task state file format, recovery patterns when an agent crashes mid-task
Error handling playbook Five failure modes (loops, context drift, partial completion, conflicting edits, silent failures) with specific fixes for each
Production checklist 15-item checklist covering architecture, state, verification, and observability — if something fails here, the system fails in production

Sample — annotated system prompt

Background worker template (excerpt)

# WHY: Workers run in isolated containers. No shared memory, no conversation

# history. If you don't put it in the prompt, they don't have it.

You are a background worker agent. You have no memory of previous sessions.

All context for this task is in this prompt. Do not assume anything that

isn't stated here.

 

# WHY: File proliferation is a common agent failure mode. Each new file adds

# maintenance burden. Agents default to "new file" when unsure — suppress this.

Edit existing files. Do not create new files unless the task explicitly requires

a new file. If you think a new file is needed, note it and stop.

 

# WHY: "I made the changes" without verification is not done. Agents must

# show their work. The check-and-report pattern surfaces issues before QA.

After each change, run the relevant tests or build command and report

the output. Do not consider a task complete until verification passes.

Error playbook — silent failures

# The failure: agent says "Done. I updated the file." File is unchanged.

 

Fix 1: Require verification output in the report.

In the worker template, add: "Your completion report must include the output of running the test/build. If you can't show verification output, you're not done."

 

Fix 2: Check from the outside.

The lead agent or QA stage reads the file directly after the worker reports completion. Don't trust the claim — check the artifact.

Pricing

$29
One-time purchase
Buy now — $29
No-questions refund if it's not what you expected