AI agents and MCP cheat sheet (TypeScript)

1. 1How agents workWhat makes a program an agent, the loop every agent runs, and how to decide whether a task needs an agent at all or a simpler workflow.6 concepts

   1. 1Agent loopA loop in your code: send the conversation and tool list to the model, run any tool calls it returns, append the results, and call the model again until it…
   2. 2Workflow vs agentIn a workflow your code decides the steps and the model fills in parts of a fixed path.
   3. 3Augmented LLMA language model given three extras by your code: tools it can call, retrieval that brings in relevant documents, and memory that carries facts between calls.
   4. 4ReAct patternThe idea, from a 2022 research paper, that a model alternates between reasoning about what to do, taking an action with a tool and observing the result before…
   5. 5Ground truth from the environmentReal results from the world, such as test output, an API response or a database row, that the agent reads after each action to judge whether it is on track.
   6. 6When not to use an agentA quick test before building: is the task multi-step and hard to specify in advance, is the outcome worth extra cost and latency, can a model do it, and can…

   6 in this chapterOpen this chapter
2. 2Tool calling from scratchHow a model asks your code to act: tool definitions, the call and result messages, tool choice, parallel calls and errors, shown with one provider API and no framework.8 concepts

   1. 1Tool definitionA name, a description and a JSON Schema for the arguments, sent with every request.
   2. 2Tool call and tool resultThe model replies with a tool call carrying an id, a tool name and JSON arguments.
   3. 3Provider differences in tool callingEvery major provider has the same loop with different field names.
   4. 4Tool choiceA request setting that lets the model decide freely (auto), forces a tool call (any or required, or one named tool), or blocks tools (none).
   5. 5Parallel tool callsOne model turn can contain several tool calls.
   6. 6Tool errors the model can act onWhen a tool fails, return a result marked as an error with a short message saying what went wrong and what to try instead, rather than throwing and crashing the…
   7. 7Writing tool descriptionsDescribe each tool as you would for a new teammate: what it does, when to use it, when not to, what it returns and any limits.
   8. 8Keeping tool results smallReturn only what the model needs to decide the next step: a few relevant fields, a page of results with a way to get more, and readable ids instead of raw…

   8 in this chapterOpen this chapter
3. Structured output and validationGetting data your code can trust from a model: JSON Schema, strict modes, validation with Pydantic or Zod, and what to do when the output is wrong.6 conceptsOpens with credits
4. Memory and stateWhat an agent remembers within one run and across runs: the context window, context engineering, compaction, long-term memory stores and explicit state.7 conceptsOpens with credits
5. Planning and multi-agent patternsThe reusable shapes for multi-step work, from simple chains to orchestrators and handoffs, and when splitting work across several agents is worth the cost.8 conceptsOpens with credits
6. Agent frameworksLangGraph, the OpenAI Agents SDK and the Claude Agent SDK: what each gives you over a hand-written loop, the core API of each, and how to choose.7 conceptsOpens with credits
7. Building and publishing MCP serversThe Model Context Protocol from the server side: roles, primitives, transports, the stateless 2026-07-28 revision, the official SDKs, testing with the Inspector, auth and the registry.8 conceptsOpens with credits
8. Evals and testing agentsHow to know an agent works before and after every change: tasks and trials, graders, trajectory checks, pass rates, model judges and a regression suite in CI.8 conceptsOpens with credits
9. Guardrails, cost and safetyKeeping agents safe and affordable: prompt injection, the lethal trifecta, least privilege, human approval, sandboxes, step and budget limits, and the cost maths of a run.8 conceptsOpens with credits
10. Observability and shippingRunning agents in production: tracing every step, OpenTelemetry conventions, durable execution, retries, streaming progress, model fallbacks and turning production traces into evals.8 conceptsOpens with credits
11. 11Glossary, A to ZEvery concept in the sheet and the smaller terms it uses, in one list.128 terms

Tool calling, MCP servers, evals.