A model proposes. The harness decides what is allowed, what is remembered and when the work is done. Give this agent a toolset, a policy and a budget, then read what happened.
3 missionsPolicy sandboxTrace + debrief
This is a simulator, not an agent. The trace is generated by a deterministic teaching model. It mirrors real harness failure modes — missing tools, approval waits, compaction loss, retry budgets and context exhaustion — without making live model calls or touching a real repository.
harness / workbenchsuccessscore 100/100
Fix the failing test
shop/checkout · The checkout total test is failing. Find the bug, make the smallest safe edit, and verify the suite. A small, well-instrumented task — the place to learn what your harness does under pressure.
P
Plan
Read the task, inspect the repository, make the smallest change, then verify it. Model: Balanced.
180 tok 900 ms
>
Search repository
Find the failing test and the cart implementation.
420 tok 2.30 s
=
Result
The result was added to context.
70 tok 150 ms
>
Read tests/cart.test.ts
Read the assertion that fails and the expected total.
780 tok 2.74 s
=
Result
The result was added to context.
70 tok 150 ms
>
Read src/cart.ts
Inspect the total calculation.
640 tok 2.57 s
=
Result
The result was added to context.
70 tok 150 ms
!
Waiting for approval
A human approved: edit · Edit src/cart.ts.
45 tok 6.50 s
>
Edit src/cart.ts
Apply the smallest fix: include the missing discount.
360 tok 2.23 s
=
Result
The file changed. The harness captured a diff for review.
70 tok 150 ms
>
Run the test suite
Execute npm test and inspect the result.
520 tok 2.42 s
=
Result
Test suite passed. The harness recorded the command, output and exit code.
70 tok 150 ms
All 14 tests pass. One file changed.
The task completed with evidence. Now look at the cost, step count and risk events.
The harness completed with a balanced toolset, a retry budget and human approval on writes. That is a solid default for real work.
Guided experiments
Same model. Different harness.
The model tier is only one decision. These experiments isolate the harness choices that turn a working demo into a system you can trust — or into an expensive incident.
01
Remove the proof.
On Fix the failing test, turn off the shell tool and run. The edit may land, but the harness cannot verify it. Read the outcome: partial, not success. Then turn the shell back on. What changed, and what did not?
02
Compress a decision away.
Choose Add input validation. Set context policy to aggressive and retries to 0. Run. Find the compaction event and the failed edit. Which constraint disappeared, and where should it have been pinned?
03
Make cheap expensive.
Switch to the fast model tier with retries at 0, then run Fix the failing test. Compare the total cost and outcome with the balanced model plus one retry. A cheap token is not the same as a cheap completed task.