Research and references
The evidence behind every pattern
28 sources, 27 verified by fetch and reading on the dates shown. Each is recorded with its publisher, its own publication date (or "undated" when it gives none), the date it was last checked, and what the library relies on it for. Sources are accepted on evidence quality, not brand. A citation is a pointer, not a licence to copy: every explanation on these pages is written in our own words.
How a claim is labelled
- evidence
- Evidence-backed recommendation from a primary source.
- convention
- Established industry convention; several sources agree.
- framework
- How one framework implements it. Version-specific.
- practitioner
- A practitioner’s observation, not independently verified.
- emerging
- An emerging technique with thin evidence.
- editorial
- Our interpretation.
How maturity is set
- established
- described the same way by several independent primary sources and in wide production use.
- emerging
- documented by at least one primary source and in use, but the shape is still settling.
- contested
- credible practitioners disagree about when, or whether, to use it.
- experimental
- promising, thinly evidenced. Treat as a hypothesis.
These rules extend the site's methodology. The interaction library promotes patterns from dated sightings across vendors; the engineering library promotes them from primary documentation and research, reviewed by an editor before publication.
Vendor guidance
- Building effective agents
The augmented LLM, the workflow/agent distinction, and five workflow patterns: prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer. Start simple; add complexity only when it measurably helps.
Cited by Single-agent tool loop, Orchestrator and workers, Stateful graph with branching and human approval
- Demystifying evals for AI agents
Grade outcomes over several trials rather than exact step sequences; combine code, model and human graders; pass@k versus pass^k; start with 20–50 tasks drawn from real failures and read the transcripts.
- Effective context engineering for AI agents
Context as a finite resource: system prompt calibration, tool design, just-in-time retrieval, compaction, structured note-taking and sub-agent architectures for long-horizon work.
- How we built our multi-agent research system
An orchestrator–worker research system: a lead agent plans and delegates to parallel subagents with their own context windows; token use explains most of the quality variance; evaluation and production hardening lessons.
Cited by Orchestrator and workers
- Claude Code: best practices for agentic coding
Practices for a single coding agent in a tool loop: CLAUDE.md context, permission allowlists, explore-plan-code-commit, test-first, subagents for verification, compaction.
- Building agents (developer learning track)
Agents as instructions, guardrails and tools; multi-agent systems are not the default; routing agents that hand off to specialists; input and output guardrails; structured outputs.
Cited by Single-agent tool loop
- A practical guide to building agents (PDF)
Single-agent systems first; the manager (tools) and decentralized (handoff) multi-agent patterns; layered guardrails; human intervention on failure thresholds and high-risk actions.
Cited by Orchestrator and workers
Framework documentation
- OpenAI Agents SDK: handoffs
A handoff is a tool call that transfers the conversation to another agent; input filters control what context travels with it.
- Microsoft Agent Framework overview
Agents and workflows as two primitives; agents for open-ended tasks, workflows for explicit, typed, multi-step orchestration.
- Microsoft Agent Framework workflows: builder and execution
Workflows as directed graphs of executors and edges with build-time type validation; a Pregel-style superstep model that gives deterministic execution and checkpointing at superstep boundaries.
- LangGraph overview
A low-level runtime for long-running stateful agents: durable execution, human-in-the-loop, short- and long-term memory; graphs of nodes and edges over shared state; it does not abstract prompts or architecture.
Cited by Single-agent tool loop, Stateful graph with branching and human approval
- LangGraph persistence
Checkpointers save graph state at every superstep into threads; enables memory, human-in-the-loop, time travel and fault tolerance.
- LangGraph interrupts (human-in-the-loop)
interrupt() pauses a graph indefinitely at a node, persists state, and resumes with a human-supplied value via Command(resume=...).
- LangGraph durable execution
Durability modes, idempotent side effects and deterministic node code so a resumed run replays safely.
- LangGraph graph API
State schemas with reducers, nodes as functions, normal and conditional edges, Send for map-reduce fan-out, subgraphs, recursion limits.
Cited by Orchestrator and workers, Stateful graph with branching and human approval
- Cloudflare Agents documentation
Stateful agents on Durable Objects, scheduling, Workflows for durable multi-step execution, human-in-the-loop with persisted state.
Standards and frameworks
- OWASP Top 10 for Agentic Applications for 2026
Listed from the brief as the agentic-application risk taxonomy; not yet read by a human from this registry.
UNVERIFIED: the site refuses automated fetches (HTTP 403). A human must open it, confirm the ten risks and the date, and update this entry before any pattern cites it as evidence.
- NIST AI Risk Management Framework (AI RMF 1.0)
Govern, Map, Measure, Manage: the risk functions the SDLC security stage and the governance notes refer to. Voluntary framework, not a certification.
- Model Context Protocol specification (2025-06-18)
Tools, resources and prompts between hosts and servers; the authorization section; the security principle that tool descriptions and results are untrusted unless the server is trusted.
Cited by Single-agent tool loop
Research papers
- ReAct: Synergizing reasoning and acting in language models
Interleaving reasoning traces and actions in a loop with environment feedback; the origin of the think-act-observe cycle.
Cited by Single-agent tool loop
- Reflexion: language agents with verbal reinforcement learning
An agent reflects on a failed attempt in natural language and stores the reflection in memory to improve the next attempt.
- Tree of Thoughts: deliberate problem solving with large language models
Branching over intermediate reasoning steps with evaluation and search; the research basis for sampling several candidates and voting.
Practitioners
- The lethal trifecta for AI agents: private data, untrusted content, and external communication
An agent that can read private data, ingest untrusted content and communicate externally can be made to exfiltrate data by prompt injection; remove one of the three capabilities.
- Don’t build multi-agents
Two principles: share full context, and remember that actions carry implicit decisions that conflict when made in parallel. Argues for single-threaded agents with context compression over parallel subagents for most tasks.
Cited by Orchestrator and workers
- How long contexts fail
Names four context failure modes (poisoning, distraction, confusion, clash) with examples from published reports.
- 12-Factor Agents: principles for building reliable LLM applications
Own your control flow, own your context window, contact humans with tool calls, small focused agents, launch/pause/resume with simple APIs, compact errors into context.
Cited by Single-agent tool loop, Stateful graph with branching and human approval
This site
- Agents & Humans: Methodology
The evidence rules the engineering library inherits: dated, sourced observations and the lifecycle vocabulary.
- Agents & Humans: the interaction pattern library
The human-facing patterns (approval, planning, interruption, recovery, memory, context, progress) each engineering pattern links to.
How the registry stays current
A research Worker checks each URL on a schedule, records the status and a content hash, and opens a proposal when something changes. A changed hash is not evidence that guidance changed; an editor reads the source before any pattern text moves. Nothing an automated worker writes reaches these pages without a human approving it in review. What's new lists every approved change.
The Agents & Humans Briefing
How humans and agents work together: the day's field notes and the patterns behind them, in one short email around noon UTC.
Free. No spam. Unsubscribe anytime.