Roadmap

Phase 1: Documentation-First Knowledge Base ✓

Restructured the repo from code-first blueprints to an architecture-first documentation system.

Delivered:

Foundations section (terminology, anatomy of an agent, pattern selection guide, plus security & safety, hallucination & grounding, evals & quality, cost & model selection, frameworks & integrations)
4 workflow patterns (prompt chaining, parallel calls, orchestrator-worker, evaluator-optimizer)
11 agent patterns (ReAct, plan & execute, tool use, memory, RAG, reflection, routing, multi-agent, event-driven, saga, human-in-the-loop)
3-tier documentation for each pattern (overview, design, implementation)
Supplementary docs for each pattern (observability, cost & latency)
Evolution bridges from workflows to agent patterns
Composition section (combination matrix, anti-compositions, reference architectures, blueprints-to-deployments, blueprint-to-spec-to-scaffold)

Add working code implementations alongside the documentation.

Delivered:

Python implementations for all 11 agent patterns (patterns/{name}/code/python/)
Python implementations for all 4 workflow patterns (workflows/{name}/code/python/)
Language-agnostic interfaces matching the pseudocode from Tier 3 docs

Remaining:

TypeScript implementations for each pattern
Formal test suites with LLM stubs for CI
Docker Compose environments for patterns requiring infrastructure (vector stores, etc.)

Expand coverage to more specialized and emerging patterns.

Delivered:

Long-horizon agents (patterns/long_horizon/) — checkpointing, resume, multi-session tasks, deep-agents shape
Agentic RAG (patterns/agentic_rag/) — query decomposition, multi-source routing, self-correcting retrieval, citation-bound answers
Guardrails modifier (modifiers/guardrails/) — layered input / tool / output policy + dual-LLM split
Sub-agents primitive (primitives/sub_agents/) — role-scoped agent instances with isolated context windows
Context Engineering foundation (foundations/context-engineering.md)

Candidates:

Delivered:

Documentation website with multi-tier navigation, search, pattern explorer, and comparison tool

Remaining:

Decisions worth recording so they don't get re-litigated. See System Design Heritage for the full reasoning.

Reliability blueprints live in agent-deployments, not here. Four reliability patterns (Circuit Breaker, Retry with Exponential Backoff, Idempotency, Distributed Tracing) are planned for the sister repo. They're operational concerns shared by every cognitive pattern, not new cognitive patterns themselves. Blueprints stays focused on the LLM-pattern layer.
Generic vs. domain-specific category axis: deferred. Every pattern in this repo is currently generic (the restaurant rebooking inside Saga is an illustrative example, not a domain pattern). Adding a domain/generic axis to metadata.json, the website, and the directory layout is premature scaffolding. Revisit when at least two domain-specific patterns exist — and consider hosting them in agent-deployments instead, since that repo already carries production-shaped, domain-bound specs.

Open an issue or start a discussion on GitHub. We prioritize based on community interest and contribution activity.