How Claude Code, Codex & co. actually work
Illustrative trace: a passing test is evidence for that case. Review the diff and relevant edge cases before accepting the change.
You ask → it tells you what to do → you open the files, run the commands, paste the errors back… and repeat.
You ask → it does the work itself — reads the files, runs the commands, checks the result → you review at the end.
The model proposes actions. The harness executes permitted tools. Together, repeating this process, they form an agent.
Who chooses an action, who runs it, and how the result comes back.
Supply relevant context, useful project rules and scoped permissions.
Inspect the changed files and run the check that defines done.
One example throughout: an invoice total that is missing VAT.
You talk to the agent. Claude Code is the harness — it runs locally with your files & terminal, and loops with the model in the cloud (which stays stateless). The agent isn't a separate thing — it's that loop.
There's no "conversation" on the model's side. Each call hands it the whole history as one flat list:
Delete the first line and the model can't answer. The "memory" lives in the request — not the model.
Context is the information supplied for this call. Tokens are the chunks used to measure it.
Optional live run: use a disposable project and scoped permissions. The next slide provides an illustrative trace if no demo is available.
In a live run or a finished transcript, point to one tool call and its result. That is one lap of the loop.
Illustrative trace. The model requests the actions; the harness runs them and returns the results.
Visible thinking may be a summary. Judge the result and the checks, not how convincing the narration sounds.
Run a coding agent beside your commands and tests.
Review and change code inside your existing development workflow.
Give a task, inspect progress, and review the returned files.
Look for clear permissions, visible changes and a way to run checks. Model and product lists are in the reference deck.
A plan for using an app or coding tool, subject to its usage limits.
Usage-based billing for integrations and automated workloads. Set a budget and monitor it.
Check the chosen product’s plan, supported access route and limits. Price per token is only one part of the cost of a completed task.
The failing test, the VAT rule and the code path that computes the total.
Old tasks, duplicate file dumps and unrelated logs that obscure those facts.
Watch for ignored requirements, repeated work and missed evidence. There is no universal percentage at which quality fails.
Start a fresh conversation when the previous task is complete.
Summarize the goal, decisions, failures and next check before continuing.
Delegate investigation and request findings with file references.
Inspect the context display as a capacity indicator. Check behavior and results before deciding what to remove.
Reuses computed state for a matching prefix. It can reduce processing cost and latency; the content is still input.
Replaces older material with a summary. It frees room, but details omitted from that summary may be lost.
For the invoice task, preserve the VAT rule, failing test, attempted changes and verification command.
Allow the project files and test commands needed for the task.
Ask before destructive actions, access changes or publishing outside the task.
For unattended experiments, restrict filesystem, network and credentials in an actual sandbox.
A separate folder or Git branch protects neither credentials nor the rest of the machine. Check the configured boundary.
Keep standing rules short. Use the instruction-file convention supported by your harness, such as CLAUDE.md or AGENTS.md.
Skill: a reusable procedure loaded when relevant. Plugin: a way to distribute a toolkit to a team.
Run the project’s invoice tests and inspect Git changes.
Fetch a permitted record through a service interface.
Expose supported issue-tracker or document tools to the agent.
Choose the narrowest useful access. A connector does not grant permission to publish or change unrelated records.
Optional reference: terminology, products and historical pricing · All talks
All figures verified September 2026. Models & prices move fast — re-check the primary sources before quoting.