LLM Agents Resources

Knowledge

• Article: "Building Effective Agents", Anthropic Engineering
  Draws the workflow against agent distinction this track is built on, then names the common composable patterns (prompt chaining, routing, parallelisation, orchestrator and workers, evaluator and optimiser) and argues for the simplest one that works. Use for: stage 1's vocabulary, and stage 16's case for not building an agent.
• Guide: "A Practical Guide to Building Agents", OpenAI
  A second vendor's framing of the same material, covering when a task warrants an agent, tool categories, orchestration for one agent against many, and guardrails. Use for: checking that a claim in stage 1 or stage 8 is about agents rather than about one provider's product.
• Article: "Writing Effective Tools for Agents", Anthropic Engineering
  Treats a tool's name, description, parameter schema and return shape as prompt surface rather than as an API contract, including how a tool's output consumes the context budget. Use for: stage 4, and for why a tool designed like a REST endpoint makes a worse tool.
• Article: "Effective Context Engineering for AI Agents", Anthropic Engineering
  Argues for treating the context window as a finite budget to be curated, and covers compaction, note-taking to external memory, and sub-agents as context isolation. Use for: stages 5 and 6.
• Paper: "ReAct: Synergizing Reasoning and Acting in Language Models", Yao et al., 2022
  Introduces interleaving reasoning traces with tool actions in one loop, rather than planning fully upfront or acting without reasoning. Use for: stage 7, and for the naming of the pattern most agent loops still implement.
• Paper: "Reflexion: Language Agents with Verbal Reinforcement Learning", Shinn et al., 2023
  Introduces self-critique as an explicit loop stage: the agent reflects on a failed attempt in natural language and carries that reflection into the retry. Use for: stage 7's reflection material, and for what self-critique costs in tokens.
• Paper: "Toolformer: Language Models Can Teach Themselves to Use Tools", Schick et al., 2023
  Shows tool use being learned by the model rather than prompted, which is the boundary between this track and llm/finetuning. Use for: stage 2's explanation of why a model emits a tool call at all.
• Docs: "Tool use with Claude", Claude Platform Docs
  One provider's own account of the tool-call wire format: the JSON schema a tool definition takes, how a tool-call request and its result are represented as messages, and parallel tool calls. Use for: stage 2's captured request and response pair, read against the OpenAI guide below as the diff.
• Guide: "Function calling", OpenAI
  A second provider's wire format for the same mechanic, including the shift from a single function call to a general tool-calling convention. Use for: stage 2, as the second provider whose format appears only as the diff from the first.
• Specification: Model Context Protocol
  The normative specification for MCP: tools, resources, prompts, sampling, the transports, and the server lifecycle. Use for: stage 9, and for settling any MCP claim from the source rather than from a client's documentation.
• Article: "Code Execution with MCP", Anthropic Engineering
  Argues that presenting tools as code an agent writes against, rather than as individual tool definitions loaded into context, changes what a large tool set costs. Use for: stage 9's tool-discovery-at-scale problem and stage 10's code-execution material.
• Article: "How We Built Our Multi-Agent Research System", Anthropic Engineering
  A production account of an orchestrator and worker system, including the token cost multiple relative to a single agent and where coordination actually broke. Use for: stage 8's argument that multi-agent buys parallelism and context isolation, at a measured price.
• Article: "Don't Build Multi-Agents", Cognition
  The opposing case, arguing that splitting work across agents fragments context and produces decisions that conflict. Use for: stage 8, read against the Anthropic account above rather than instead of it.
• Article: "The Lethal Trifecta for AI Agents", Simon Willison
  Names the combination that makes an agent exploitable: access to private data, exposure to untrusted content, and the ability to communicate externally. Use for: stage 14's central framing, and for why prompt instructions are not a defence.
• Reference: "OWASP Top 10 for LLM Applications", OWASP GenAI Security Project
  The consensus catalogue of LLM application risks, including prompt injection, excessive agency and supply chain exposure, each with mitigations. Use for: stage 14's checklist, and for the vocabulary a security reviewer will already be using.
• Specification: "Semantic Conventions for Generative AI", OpenTelemetry
  Vendor-neutral conventions for the spans and attributes an agent run emits, which is what makes one tracing backend's trajectory view readable in another. Use for: stage 12, as the provider-neutral answer to what a trace should contain.
• Paper: "tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains", Yao et al., 2024
  Evaluates agents on multi-turn tool use against a simulated user and a domain policy, and introduces a repeated-trial reliability measure rather than single-trial success. Use for: stage 13's distinction between succeeding once and succeeding reliably.
• Paper: "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?", Jimenez et al., 2023
  Evaluates agents on real repository issues with the project's own tests as the outcome check. Use for: stage 13's example of an outcome-graded benchmark where the trajectory is unconstrained.
• Docs: "Computer use tool", Claude Platform Docs
  Describes the screenshot-and-action loop a computer-use agent runs: an observation is a screenshot rather than a tool result, and an action is a screen coordinate or a key press rather than a function argument. Use for: stage 11's grounding-an-action-to-a-coordinate material, and for why this loop needs a different reliability story than a tool-calling one.
• Paper: "WebArena: A Realistic Web Environment for Building Autonomous Agents", Zhou et al., 2023
  A self-hosted web environment with functional correctness checks, built for agents that act through a browser. Use for: stage 11, as the standard against which browser-agent reliability claims are made.
• Paper: "GAIA: A Benchmark for General AI Assistants", Mialon et al., 2023
  Questions that are easy for a person and hard for an agent, requiring multi-step tool use with a single unambiguous answer. Use for: stage 13, as the counterweight to benchmarks a model can pass without acting.
• Docs: "Prompt caching", Claude Platform Docs
  Caching a prompt prefix to cut repeated cost and latency, including why an agent's stable tool definitions and system prompt are exactly the prefix worth caching and what silently invalidates the cache. Use for: stage 15's token-amplification and cost material.

Gaps

• No provider-neutral reference exists for tool-calling wire formats. Each provider documents only its own, so the captured request and response pairs this track shows have to be taken from provider documentation and reconciled by hand.
• Durable and resumable agent execution is documented mainly inside individual frameworks, which is neither provider-neutral nor stable enough to list yet. Stage 15 needs a source that is neither.
• Computer use and browser agents move faster than anything written about them. The listed benchmark papers cover how to evaluate one, not current practice for building one, so stage 11 is the thinnest-sourced part of the arc.
• Nothing listed yet covers agent memory across runs as a designed system rather than as a product feature. Stage 6 needs one.