Skip to content

Latest commit

 

History

History
105 lines (77 loc) · 5.28 KB

File metadata and controls

105 lines (77 loc) · 5.28 KB

Blueprint Map

Full mapping between agent-deployments recipes and agent-blueprints patterns.

Authoritative machine-readable view: catalog.yaml. This document is the prose-friendly companion — it adds context and rationale (per-recipe pattern justification, composition/ notes, visual mapping) that the catalog can't carry. When the two disagree, the catalog wins; open a PR to fix this file.

agent-blueprints teaches architecture at three tiers: Overview, Design, Implementation. Each blueprint below links to the relevant pattern pages.

Recipe-to-pattern mapping

Recipe Primary pattern(s) Blueprint links
customer-support-triage Routing + Tool Use patterns/routing/ (overview, design, implementation) · patterns/tool-use/ · foundations/choosing-a-pattern.md
docs-rag-qa RAG patterns/rag/ (overview, design, implementation)
research-assistant ReAct + Tool Use patterns/react/ (overview, design, implementation) · patterns/tool-use/ · patterns/react/evolution.md
content-pipeline Prompt Chaining + Evaluator-Optimizer workflows/prompt-chaining/ · workflows/evaluator-optimizer/ · composition/
code-review-agent Plan & Execute + Reflection patterns/plan-and-execute/ · patterns/reflection/ · patterns/plan-and-execute/evolution.md
ops-crew Multi-Agent (flat) patterns/multi-agent/ (flat variant)
parallel-enricher Parallel Calls workflows/parallel-calls/
memory-assistant Memory patterns/memory/
hierarchical-agent Multi-Agent (hierarchical) patterns/multi-agent/ (hierarchical variant)
restaurant-rebooking Event-Driven + Multi-Agent (flat) patterns/event-driven/ · patterns/multi-agent/ (flat variant) — first recipe to declare capabilities: end-to-end
claude-code-subagent ReAct + Routing/Tool Use patterns/react/ (overview, design, implementation) · patterns/routing/ · patterns/tool-use/ — host CLI delegating a research turn to a Claude Code-style subagent via the Claude Agent SDK

How to read this

Each blueprint's README opens with a Blueprint Map block that links directly to the relevant agent-blueprints pages. For example:

## Blueprint Map

- Overview: agent-blueprints/patterns/react/overview.md
- Design: agent-blueprints/patterns/react/design.md
- Implementation: agent-blueprints/patterns/react/implementation.md
- Evolution: agent-blueprints/patterns/react/evolution.md

Visual mapping

agent-blueprints                          agent-deployments
─────────────────                         ──────────────────

foundations/choosing-a-pattern.md    ──►   customer-support-triage
                                           (why Routing, not ReAct)

patterns/routing/                    ──►   customer-support-triage, claude-code-subagent
patterns/tool-use/                   ──►   customer-support-triage, research-assistant, claude-code-subagent

patterns/rag/                        ──►   docs-rag-qa

patterns/react/                      ──►   research-assistant, claude-code-subagent

patterns/plan-and-execute/           ──►   code-review-agent
patterns/reflection/                 ──►   code-review-agent

patterns/multi-agent/  (flat)        ──►   ops-crew, restaurant-rebooking
patterns/multi-agent/  (hierarchical)──►   hierarchical-agent

patterns/event-driven/               ──►   restaurant-rebooking

patterns/memory/                     ──►   memory-assistant

workflows/prompt-chaining/           ──►   content-pipeline
workflows/evaluator-optimizer/       ──►   content-pipeline
workflows/parallel-calls/            ──►   parallel-enricher

composition/                         ──►   content-pipeline (headline example)
                                           hierarchical-agent

Capabilities layer

Layered orthogonally on top of the recipe → pattern mapping, docs/capabilities/ describes the provisioning contracts consumed by agent-scaffold up (≥ v0.3):

docs/capabilities/
  vector_db/{qdrant, chroma, pgvector}
  cache/{redis}
  relational/{postgres}
  queue/{kafka, redis-streams}
  obs/{langsmith, langfuse, grafana-stack}
  eval/{promptfoo}
  frontend/{nextjs-chat, streamlit}
  host/{vercel, railway, fly}

A recipe opts in by declaring capabilities: [...] in its frontmatter. The scaffold's resolver matches each id against the catalog and threads the resolved bodies through context assembly + orchestrator bootstrap steps. See docs/capabilities/README.md for the schema and authoring guide.

Contributing

When adding a new blueprint, you must:

  1. Identify the corresponding pattern(s) in agent-blueprints
  2. Add the Composes section to the recipe linking back to the pattern
  3. Update this file with the new mapping
  4. Coordinate a PR to agent-blueprints adding a cross-link back to this recipe
  5. If the recipe needs new infra not in docs/capabilities/, add the capability file alongside the recipe PR (or in a sibling PR)