1 min lesson
What 'agent-harness performance' means here
Use "The agent harness is the loop wrapped around the model" to say what a strong answer must include.
Step 1 of 3
The agent harness is the loop wrapped around the model: it gathers context, calls tools, applies edits, coordinates across files and streams results back. The model writes the tokens. The harness is everything that turns those tokens into a working change in your repo.
Every stage of that loop is a place reliability can break and most of them have nothing to do with model quality. A perfect model still produces a bad experience if the edit fails to apply, a tool call times out or the run hangs halfway through a multi-file change.
Learn more
Full explanation
The Harness, Layer by Layer
Interactive diagram. Tab through its regions; each focused region shows its detail in the panel below.
Each layer wraps the model and is a place reliability can break - most failures live here, not in the tokens.
The distinction you must hold crisply is model quality versus harness reliability. Model quality is whether the LLM's output is correct and useful. Harness reliability is whether the loop executed quickly and correctly regardless of the output. This seat leans heavily on the second and the hard analytical move is separating the two when a run goes wrong.
- Symptom a user reports
- Agent gave a wrong answer
- Likely model issue
- Reasoning or knowledge gap in the model
- Likely harness issue
- Wrong or missing context retrieved into the prompt
- Symptom a user reports
- Agent run never finished
- Likely model issue
- Rarely the model
- Likely harness issue
- Timeout, hung tool call or stalled stream
- Symptom a user reports
- Edit didn't show up in the file
- Likely model issue
- Almost never the model
- Likely harness issue
- Edit-application failure or a dropped diff
- Symptom a user reports
- Completion felt sluggish
- Likely model issue
- Larger/slower model variant
- Likely harness issue
- Latency in retrieval, queueing or serving
| Symptom a user reports | Likely model issue | Likely harness issue |
|---|---|---|
| Agent gave a wrong answer | Reasoning or knowledge gap in the model | Wrong or missing context retrieved into the prompt |
| Agent run never finished | Rarely the model | Timeout, hung tool call or stalled stream |
| Edit didn't show up in the file | Almost never the model | Edit-application failure or a dropped diff |
| Completion felt sluggish | Larger/slower model variant | Latency in retrieval, queueing or serving |
The same complaint splits into model vs. harness causes - and your metrics have to tell them apart.