‘42 Agent Architecture Patterns’ — new collection
Carlos E. Perez posted a roundup titled “42 Agent Architecture Patterns,” covering things like skill repositories and intent engineering with a demo video—useful reference material for building production agent systems. The collection is a practical taxonomy for agent‑centric system design. (x.com)
Perez frames the pattern set as an Alexandrian-style pattern language targeted at Claude-code development and explicitly argues that prompting alone is insufficient, urging deterministic structural constraints for reliable agents. (youtube.com) The material groups patterns into structural foundations (flat skill catalog, skill-directory anatomy, YAML front matter), context engineering (three-tier knowledge loading, write-to-file, source‑attributed state), harness engineering (explicit I/O contracts, pipeline composition, checkpointing), and intent engineering (identity‑first rules, constitutions, philosophy injection). (youtube.com) Practical, demoed recommendations include file-based data passing between components, clear writer/evaluator role separation, bounded correction loops, and a deterministic-versus-creative split to limit nondeterminism in production runs. (youtube.com) The video presents applied Claude-code walkthroughs and a recorded demo that ties the pattern prescriptions to concrete code and harness examples on the Intuition Machine channel. (youtube.com) Intuition Machine’s creator markets related products—an advertised Deep Learning AI Playbook, paid workshops, and a pattern-language bundle—on a Gumroad storefront that links this pattern work to paid teaching materials. (intuitionmachine.gumroad.com) The presentation explicitly names operational tensions—context richness vs. context-window limits, autonomy vs. safety, and speed vs. correctness—and recommends adversarial testing and file-based checkpoints as scaling controls for long‑running agent systems. (youtube.com)
