Evaluating Agent Patterns Catalog as a Source¶
A CC BY 4.0 catalog of 421 agentic patterns — useful as a citation index when co-cited with primary sources, not as an MCP server.
Agent Patterns Catalog (agentpatternscatalog.org) is a CC BY 4.0 reference catalog for agentic-systems design patterns, maintained by independent operator Marco Nissen and grounded in the open agentpatternscatalog/patterns GitHub repository. It earns inclusion in our research source list with explicit guard-rails — cite by URL with attribution, co-cite the linked primary source on the same claim, do not wire its hosted MCP server, and distinguish it from Liu et al.'s academic catalog of the same name (arxiv 2405.10467).
What the catalog is¶
The catalog frames itself as "A Pattern Language for Agentic Systems" — a nod to Christopher Alexander's 1977 book 'A Pattern Language' (catalog home). At the time of evaluation it advertises 421 patterns, 161 compositions, 49 methodologies, 14 books, and 90 trainings, surfaced through a typed-relation graph: each pattern declares specialises, complements, or alternative-to links to its neighbors.
Each pattern page follows a Gang of Four / POSA convention: Context, Problem, Forces, When to use, When not to use, Example, Solution ("Therefore: …"), What it gives you, What it costs you, References, and Provenance. The Toolformer pattern page stamps a GitHub commit hash, added date, and last-updated date — provenance is per-page, not site-wide.
Machine-readable surfaces are first-class: patterns.json and compositions.json are CORS-open with a 5-minute cache, and an llms-full.txt dump targets LLM consumption.
Why it works as a citation source¶
The catalog reduces our research-fetch cost on pattern-shaped topics by serving as a low-cost index over the primary sources we would cite anyway. Each pattern page links out to the underlying primary references — Toolformer to arxiv 2302.04761, Agent Skills to Anthropic's agent-skills documentation — so it functions as a pre-structured starting point, not as the authority on the claim itself. This is the same mechanism that justifies our existing inclusion of nibzard/awesome-agentic-patterns in SOURCES.md: the index is the discovery surface, the primary source is cited for the claim.
The catalog's CC BY 4.0 license and explicit attribution string ("Agent Patterns Catalog — agentpatternscatalog.org") match the link-and-attribute posture our editorial inclusion criteria already require — no per-page negotiation needed. Operator identity is named (Marco Nissen, marco-nissen.com) and per-pattern provenance carries a verifiable GitHub commit hash.
The structural match also helps. Our anti-patterns/ and fallacies/ sections already follow a context-pressure-consequence shape; the catalog's Prompt Bloat anti-pattern maps onto it without translation, making cross-checks against an outside catalog cheap.
When this backfires¶
Three failure modes turn the catalog from a useful index into a liability.
Wholesale reproduction. CC BY 4.0 permits redistribution with attribution, but copying pattern bodies into our pages dilutes our editorial voice and produces duplicate content — a GEO and AEO negative, and a violation of our arxiv-style ToU posture. The acceptable form is a short quote plus a link, never paragraph-scale reproduction.
Treating it as authoritative on disputed claims. The GitHub repository shows commits authored solely by Marco Nissen, with 2 stars and 0 forks as of this review (June 2026). This is normal for a young single-maintainer catalog, but it means peer review is light. On topics where the academic literature is unsettled (multi-agent coordination, memory architectures, evaluation methodology) the catalog is one signal among many, not the closing argument. Co-cite a primary source on the same claim or do not make the claim.
Wiring the hosted MCP server. The catalog exposes a hosted Model Context Protocol endpoint at https://mcp.agentpatternscatalog.org/mcp with tools like find_pattern, recommend_recipe, and pattern_for_symptom (Use page). Adding it to a sub-agent that also holds repo-read access would close the lethal trifecta on that principal — private data, untrusted content, egress — exactly the scenario our AGENTS.md MCP-onboarding gate exists to prevent. This evaluation is a citation-source decision, not an MCP-wiring decision — the catalog goes in SOURCES.md, not .mcp.json.
A fourth, smaller risk is name collision: search engines conflate this catalog with Liu et al.'s peer-reviewed Agent Design Pattern Catalogue (CSIRO Data61, 18 architectural patterns). They are distinct artifacts with overlapping vocabulary. Citations must disambiguate.
Recommendation¶
Use it as a source, with guard-rails. Add it to SOURCES.md in the same form as the existing nibzard/awesome-agentic-patterns entry: cite individual patterns by pattern-page URL with the recommended attribution string, never reproduce pattern bodies, and always co-cite the pattern's linked primary source on the same claim — the catalog is the index, not the authority. Do not wire the MCP server without first running agent-readiness-audit-lethal-trifecta, and disambiguate the site catalog from Liu et al.'s CSIRO catalog in any citation that names "Agent Patterns Catalog."
The strongest near-term uses: scanning the anti-patterns book for failure modes our anti-patterns/ and fallacies/ sections have not yet named, and mining the typed-relation graph for sibling-link suggestions when authoring new pages in patterns/ or agent-design/.
Key Takeaways¶
- The catalog is CC BY 4.0, single-maintainer, GoF-shaped, and grounded in a public GitHub repository — adequate for citation, light on peer review.
- Use it as an index over primary sources, not as the authority on a claim. Always co-cite the underlying reference.
- Do not wire its hosted MCP server without running the lethal-trifecta audit; the citation decision and the MCP decision are separate.
- Add to
SOURCES.mdwith the same posture used fornibzard/awesome-agentic-patterns: cited individually, summarised in our own words, never reproduced wholesale.