Agent Atlas

Index

The Atlas

Every idea on the plates, explained at three depths, with the sources behind it.

Start here

New to all of this? Take these twelve in order. Each is a short page, and most point to a plate where you can try it.

  1. Language model A program trained on vast amounts of text to predict what comes next. Text goes in, text comes out. Try it
  2. Tokens The pieces text is cut into before a model reads it. Size, limits and prices are counted in them. Try it
  3. Context window Everything the model can see on one turn, with a hard size limit measured in tokens. Try it
  4. The agent loop Call the model, carry out what it asks for, add the result, and repeat until it answers in words. Try it
  5. Tool use The model replies with a structured request naming a tool and its inputs; the harness decides whether to run it. Try it
  6. The terminal The text interface most coding agents work through. Try it
  7. Instruction files AGENTS.md, CLAUDE.md and their relatives: text the harness loads before your request. Try it
  8. Agent memory A new session starts with an empty context. What an agent like Claude Code remembers comes back from files. Try it
  9. Context rot Models get less reliable as their context grows, and use the middle of it least. Try it
  10. Sampling A model scores every possible next token. A sampler then picks one, using settings such as temperature and top-p. Try it
  11. Model and effort Choosing a model and how much effort it spends. Two Claude Code builders publish when they go low, medium or high. Try it
  12. Verification Checking the result before claiming success. Try it

Understanding agents

  • The agent loop Call the model, carry out what it asks for, add the result, and repeat until it answers in words.Plate I
  • Harness The program around the model that builds context, runs tools, enforces rules and keeps state between sessions.Plate I, Plate II
  • Instruction files AGENTS.md, CLAUDE.md and their relatives: text the harness loads before your request.Plate II, Plate III
  • Asking well Your request is most of what the model knows about your task. Say what you want, why, and what done looks like.Plate I
  • Tool use The model replies with a structured request naming a tool and its inputs; the harness decides whether to run it.Plate I
  • Agent memory A new session starts with an empty context. What an agent like Claude Code remembers comes back from files.Plate II, Plate IV
  • Hooks Scripts the harness runs at set moments, such as before every tool call, that can block an action and say why.Plate II
  • Permissions Rules the harness applies before a tool runs, outside the model's control.Plate I, Plate II
  • Skills Packaged instructions for one kind of task, loaded only when a task needs them.Plate II
  • Verification Checking the result before claiming success.Plate I
  • Model Context Protocol A standard way to plug extra tools and data into a harness.Plate II
  • Plugins One installable bundle of skills, subagents, hooks and tool servers.Reference
  • Sandboxing A wall the operating system puts around what an agent's commands can touch, stronger than any instruction in a prompt.Plate II
  • Subagents A helper agent that works in a separate context and returns its results to the main agent.Reference

Understanding models

  • Context window Everything the model can see on one turn, with a hard size limit measured in tokens.Plate I
  • Language model A program trained on vast amounts of text to predict what comes next. Text goes in, text comes out.Plate I, Plate V
  • Tokens The pieces text is cut into before a model reads it. Size, limits and prices are counted in them.Plate I, Plate V
  • Compaction Shrinking older context into a summary to make room. Details can be lost.Plate I
  • Context rot Models get less reliable as their context grows, and use the middle of it least.Plate IV
  • Cost and latency Why long sessions get slower and more expensive, and what caching changes.Plate I
  • A short history The papers and releases behind today's models, in order, each with its date.Reference
  • Sampling A model scores every possible next token. A sampler then picks one, using settings such as temperature and top-p.Plate V
  • Post-training Pretraining learns patterns in text. Post-training, with examples and feedback, makes a model better at following instructions.Reference
  • Quantization Storing each weight in fewer bits to save memory, with a possible loss in accuracy.Plate V
  • Scaling laws As models, data and training compute grow, next-token prediction loss falls in a predictable way.Reference
  • Transformer The architecture most of today's language models build on: stacked layers that mix tokens together through attention.Reference
  • Mixture of experts A model with sparse expert layers, where a gate picks a few experts for each input. Active and total parameter counts differ.Reference
  • Serving Running a trained model to answer requests: the KV cache, memory paging, and computing several tokens at once.Plate V

Computer fluency

  • Files and folders Where agents read and write. A path names a file by the folders that lead to it.Plate III
  • The terminal The text interface most coding agents work through.Plate III
  • Git Version history that makes agent work safe to try and easy to undo.Reference
  • Package managers Tools that fetch the code your project depends on, and record exactly which versions you got.Reference
  • Private network Your own machines linked so they can reach each other from anywhere, with the traffic between them encrypted.Plate VI
  • SSH A way to log into another computer and run commands on it, with everything encrypted on the way.Plate VI
  • Processes and signals Every running program is a process. Signals are how you ask one to stop, or make it stop.Reference

Building real systems

  • APIs and HTTP How one program asks another for something over the web. Every agent talks to its model this way.Plate VI
  • Logs The running record of what a program did, and the first place to look when something goes wrong.Reference
  • Model and effort Choosing a model and how much effort it spends. Two Claude Code builders publish when they go low, medium or high.Reference
  • Config and secrets Settings that change between machines, and keys that must never be in the code an agent can read and commit.Reference
  • Tests and CI Checks that run automatically on every change, so a broken commit is caught before anyone ships it.Plate II
  • Deploys Putting a new version where people can use it. A good deploy is one command and easy to undo.Plate II
  • Evals Repeatable tests for an AI system, so you can tell whether a change made it better or worse.Reference
  • Worktrees Several checkouts of one repository side by side, so several agents can work on different branches at once.Reference

The frontier

  • Open weights Some labs publish the model file itself, so anyone can run it. A weights release does not by itself include the data or code.Reference
  • Chips and data centers Models run on accelerators with fast memory, packed into buildings measured in megawatts.Plate V