Introduction

Level of harness

1. Prompt management
2. Tool orchestrator
3. Memory management
4. Error handling
5. Input/output formatting
6. Guardrails

How to implement

Flows

1. You —> give task
2. Agent reads harness files
3. Agent executes, harness governs every step:
4. Instructions: what to do, in what order
5. Scope: one feature at a time, no overreach
6. State: progress log, feature list, git history
7. Verification: tests, lint, type-check, smoke runs
8. Lifecycle: init at start, clean state at end
9. Agent stops only when verification passes

Session 1: agent reads instructions agent runs init.sh agent works on one feature agent verifies before claiming done agent updates progress log agent commits clean state

Session 2: agent reads progress log agent picks up exactly where it left off agent continues the unfinished feature you review, not rescue

Result: agent does the work, you verify the result

PROJECT ROOT
  ├── AGENTS.md              <-- the agent's operating manual
  ├── init.sh                <-- runs install + verify + start
  ├── feature_list.json      <-- what features exist, which are done
  ├── progress.md            <-- what happened each session
  └── src/                   <-- actual code