Open Source

AAMAD Framework: Multi-Agent Development Methodology (PyPI Published)

Framework Creator & Maintainer|

Ongoing

80% (pre-AAMAD: 6/10 off-spec, post: 1/10)

Agent Drift Reduction

Published to PyPI, 3 production deployments

Framework Adoption

13 test plans, 89 test cases validating contracts

Test Coverage

3 adapters (CrewAI, LangGraph, Prompt-based)

Adapter Compatibility

The Challenge

AI-assisted development with multiple agents (Product Manager, Backend Dev, Frontend Dev, QA, DevOps) lacks standardized contracts, leading to context drift, missed requirements, and inconsistent quality. Needed a framework with strict persona contracts, reproducibility, provenance tracking, and adapter abstraction for CrewAI, LangGraph, or prompt-based execution.

The Approach

Created AAMAD (AI-Assisted Multi-Agent Development) with 5 core principles: (1) Single-responsibility personas with 11-section contracts (Identity, Inputs, Outputs, Workflow, Quality Gates, Failure Policy, etc.), (2) Context-first engineering — all outputs trace to PRD/SAD/SFS, (3) Reproducibility via temp-write-atomic-replace and audit trails, (4) Adapter abstraction for CrewAI/LangGraph/prompt execution, (5) Deterministic outputs with quality gates and self-checks. Published to PyPI with CLI, validated across 13 test plans and 89 test cases.

Key Learnings

Strict persona contracts prevent 80% of AI agent drift and hallucinations
Adapter abstraction enables framework-agnostic development — switch CrewAI to LangGraph without changing personas
FAILURE: Early version had 7-section contracts — too minimal. Expanded to 11-section standard after validation kept failing.
Open-sourcing frameworks forces you to generalize and document — improves your own thinking