Reference

Map of the era

One path from what the model sees to the names people use in 2026. Walk it in order. Later layers are ways of spending a token budget on purpose — not magic.

Anthropic’s working definition of an agent is an LLM using tools in a loop. LangChain’s point is that a loop is already a tiny graph, and a harness is the software wrapped around that loop. The “era” is those layers stacking, not a new kind of intelligence each month.

1. 01
   Tokens and next-token prediction

   The model does not read English. It reads a sequence of token IDs and emits the next one.
2. 02
   The context window is RAM

   One finite slot for system, history, tools, and output. Overflow vs rot vs the U-curve.
3. 03
   The API call and sampling

   Messages as a play script. Pack · read · pick. Temperature = safe vs spicy.
4. 04
   Look it up, then ask (RAG)

   Find a few pages, paste them into the script, then pick. Skip the shelf when you already have it.
5. 05
   The model writes a slip (tools + MCP)

   Propose · run · paste · pick. The model has a pen, not a thermometer. MCP is USB-C for kitchens.
6. 06
   Two architectures (workflow vs agent)

   You write the program. Workflow = your path. Agent = your loop; the model picks the next tool.
7. 07
   Harnesses

   The box around the loop: allow, run, stop, show. Cursor is an agent inside a harness.
8. 08
   Skills (progressive disclosure)

   A folder, not a second model. Catalog in RAM; load the chapter only when this task needs it.
9. 09
   Context engineering (evict · compact · note)

   Curate the attention budget. Evict dumps, compact into a summary, write a note to disk.
10. 10
    Just-in-time files

    Keep a path, not a page. Glob / grep / read this turn. The corpus stays on disk.
11. 11
    Sub-agents

    Give a helper a clean window; take back a short summary, not the search dump.
12. 12
    Loop engineering

    Design the cycle that prompts the agent: verify, stop, schedule — don’t type every turn.
13. 13
    Graph engineering

    Encode your knowledge as nodes and edges when a single loop is the wrong shape.
14. 14
    Evals and production judgment

    Measure, then add complexity only when it moves a number you care about.
15. 15
    Development flow (team floor)

    Write guidelines, guardrails, and named jobs in the repo so humans and agents share one floor.
16. 16
    Connecting the stack

    Name the failure, then climb: context, loop, or graph. Harness and floor wrap every rung.
17. 17
    Plugins

    A plugin is how Guide, Guard, Job, and MCP arrive — not a new model.

Advanced path

1. 18
   Production traces

   One request id across HTTP, agent spans, tools, and the world.
2. 19
   Production evals

   Capability vs regression; graduate the hill; calibrate the judge.
3. 20
   Agent security

   The prompt is not the wall: allowlist, untrusted text, secrets out.
4. 21
   Build track 1 · Write the loop

   The tool loop in TypeScript against a scripted model. No API key.
5. 22
   Software factory 1 · The lights go out last

   Seed · validate · feed back. Name what replaced the human reading the diff.
6. 23
   Software factory 2 · A check the builder cannot see

   Hold out · twin · score. Build the check in TypeScript and catch two bad builds.
7. 24
   Software factory 3 · A repo agents can read

   Record · map · rule. Put each fix in the right place; write a lint whose message is the fix.
8. 25
   Software factory 4 · The ramp

   Shadow · release · keep. Stop reading one class of change at a time; cap the bill; know where a human stays.

Sources for the stack of names: Building effective agents; Effective context engineering; Deep Agents vs LangChain vs LangGraph; 3 Years of Graph Engineering; Demystifying evals for AI agents; CLAUDE.md; OpenTelemetry traces. Full list in RESOURCES.md.