/ interactive
Interactive
Practice, not just reading. Test yourself against the core ideas, then build the harness layer by layer through the guided projects.
The build-along track
Read three real harnesses, then build your own one layer at a time — the loop, the tools, the context engine, durability, orchestration. Study the case studies alongside your build.
Quiz yourself
0 / 12Twelve questions spanning the loop, tools, context, durability and orchestration. Pick an answer to see the explanation.
Q1
What best describes the difference between a model and a harness?
A model does transactional inference; the harness is everything around it that turns that into a working agent.
Q2
At its core, the agent loop is…
The whole of agency is that while-loop: the model decides, the harness executes, and the result goes back in.
Q3
In a harness, what actually holds the agent's state between turns?
The array of typed messages IS the agent's memory-so-far; each turn appends to it.
Q4
A tool, as the model sees it, is essentially…
The schema is the contract between the model and your code — get its shape right and the model calls it correctly.
Q5
Why does a good harness ask before running `rm -rf` or a destructive command?
Approval modes / allow-lists are the human-in-the-loop gate that keeps an autonomous agent safe.
Q6
What problem does sandboxing a harness solve?
An agent runs arbitrary commands; the sandbox limits what a bad call can touch.
Q7
Compaction in a harness means…
The context window is finite; compaction keeps a 200-turn session coherent without blowing the budget.
Q8
The CLAUDE.md pattern is about…
It's durable project facts, reloaded each run, so the agent doesn't relearn the codebase every time.
Q9
Durable execution / checkpointing gives a harness…
Every model turn and tool call is a persisted step; on restart, replay skips the work already done.
Q10
When should a harness spawn a sub-agent?
Sub-agents partition work that won't fit one context; the parent gets the distilled result, not the whole transcript.
Q11
Why is the system prompt described as 'infrastructure, not prose'?
It's the operating manual the harness builds programmatically, not a static personality blurb.
Q12
pi (pi.dev) is notable among harnesses mainly for…
pi is the proof a real harness can be small: a tiny core plus extensions, any model behind models.json.