Agentic Engineering Patterns Library

Most agent systems fail the same handful of ways: the single agent given twelve tools that picks the wrong one, the subagent team that was really a sequential job, the loop with no exit, the memory that poisons the next run, the agent that claims it's done and isn't. The patterns that avoid them exist — they're just scattered across blog posts, papers, and other people's postmortems. This is the reference: what holds up in production, when each pattern applies, when each is the wrong answer, and what it costs. It ships as seven reference files. A pattern index maps symptoms and requirements to the right family and sets the decision order, because getting decomposition wrong cannot be fixed downstream. Orchestration covers single agent, pipeline, router, planner-executor, subagent teams, supervisor, and iterative refinement — with the for-loop test that kills most fake multi-agent designs. Context covers progressive disclosure, retrieval, compaction and what it silently loses, scratchpads, isolation, memory hygiene, and the rent every token pays on every call. Tool design covers granularity, descriptions as the real interface, argument design that doesn't force the model to invent values, read/write separation, authorization gates, and error returns that teach instead of triggering blind retries. Failure and recovery covers budgets and stop conditions, loop detection on state rather than repetition, retry discrimination, checkpointing, visible degradation, the defined give-up, and idempotency for a world where everything runs twice. Oversight and evaluation covers approval gates versus approval fatigue, review sampling weighted to risk, verification over self-reports, golden sets, adversarial cases, judge bias and its mitigations, and trajectory evaluation. A final file covers composition — which combinations reinforce, which quietly fight — and works a full support-triage agent design end to end, including the patterns it rejects and why. It is a reference library and design aid, not a framework, runtime, or code generator: it doesn't install anything, run your agent, or write your tool implementations, and patterns are judgment aids rather than guarantees. Works with Claude Code, Cursor, Codex CLI, Gemini CLI, and any SKILL.md agent.