conformance-fleet

A conformance-aware fleet of agents for OpenClaw, powered by Amazon Bedrock.

14 agents. Five conformance gates. One rule above all: the orchestrator never writes code.

This repository contains the bootstrap prompts described in the companion blog post: How I Built a Team of AI Agents I Can Finally Trust (and the Conformance System That Made It Possible)

What's in this repo

File	Description
main-agent.md	Bootstrap prompt for the main orchestrator, including the delegation-first mandate, five conformance gates, routing matrix, and TPM planning mode
sub-agents.md	Bootstrap prompts for all 13 sub-agents, each with role-specific mandates, conformance responsibilities, and output contracts. Also contains the workspace and model tier configuration.

The fleet

Tier	Agents	Model	Why
Orchestration and Governance	main-orchestrator, intent-guardian, experiment-designer, team-manager, principal-sde	Claude Opus 4.7	Blocking decisions that cascade silently if wrong
Implementation and Domain	sde-ii, security-engineer, product-manager, qa-engineer, genai-engineer, analysis-engineer	Claude Sonnet 4.6	Execution and domain review; hard reasoning done upstream
Efficiency Roles	ux-engineer, devops-engineer, finops-engineer	Claude Haiku 4.5	Checklist-driven, pattern-matched reviews

Requirements

• An AWS account with access to Anthropic Claude models on Amazon Bedrock
• OpenClaw installed and configured
• AWS credentials configured (SSO, instance profile, or environment variables)

Setup

1. Bootstrap the main agent

Open a new OpenClaw session and ask it to bootstrap itself from the main agent file:

bootstrap from https://raw.githubusercontent.com/orenGH/conformance-fleet/main/main-agent.md

2. Bootstrap the sub-agents

In the same session, ask it to bootstrap the sub-agents:

bootstrap from https://raw.githubusercontent.com/orenGH/conformance-fleet/main/sub-agents.md

3. Start building

Give the orchestrator a task. It will identify the required roles, ensure an Intent Brief exists, run the conformance gates, delegate implementation, and close only when all gates pass.

You don't need all 14 agents on day one

The minimum viable starting point:

1. One orchestrator with the delegation-first mandate
2. A principal SDE + Implementation Engineer split (design separated from execution)
3. An intent guardian with a formal Intent Brief
4. A FINGERPRINT.json baseline for any work involving GenAI or benchmarks

Add roles as you hit the gaps they were built to fill. The blog post walks through each gap and the failure that motivated the role.

License

MIT-0

---

The views expressed in this repository are those of the author and do not represent the views of Amazon Web Services.