RESEARCH · HARNESS LAB

Harness scorecard

What AI coding agent harnesses record on your machine, and what they send off it, compared with what their vendors say. One row per question, one column per harness.

No results are published yet. Where a version is shown, testing is done and results will appear once any vendor has had time to reply. The first results will follow in October. Each cell will show what one version did on the date we tested it. We re-test when new versions ship. This is not a ranking and we do not certify products. Read how we test.

Question	Claude Code	Codex	Amp	Auggie	Kilo	Qwen Code	Kimi CLI	Cline	OpenCode	Pi	Goose	Aider	Grok CLI	DeepSeek Harness	ZCode (GLM)
Version tested	2.1.283	0.157.1	0.0.1790380893	0.36.0 (static only)	7.8.1	0.24.6	Kimi Code 2.1.1	3.0.65	1.18.32	0.87.1	1.52.0	0.86.2	1.0.41	0.1.5-rc.3 (latest) 0.1.7-rc.2 (next)	Queued
Date tested	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	26 Sep 2026	Queued
What it records
What it records on the machine, and where	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Prompt, tool calls, tool output, model and time recorded	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Approvals recorded, and by whom	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Record protected from reading by others	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Record protected from change by the agent (tamper-evident)	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
How long the record is kept by default	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
What it sends
Traffic besides the model call, and to whom	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Extra fields or identifiers added to model calls	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Fake secret left the machine	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Documented opt-outs work as stated	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
Resume or feedback commands upload unexpected data	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending
What the vendor says
Observed behaviour matches vendor privacy statement	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending	Pending

Legend

Pending: not yet tested.
Yes or No: what we observed, with a link to the evidence note.
Partly: true in some runs or settings but not others. The note explains which.
Unknown: we could not observe it. Anything that happens on a vendor's servers, such as retention or training use, is unknown unless the vendor has stated it.
Not applicable: the harness has no such feature.
Awaiting vendor reply: we have shared a finding privately and the reply period of 14 to 21 days is still running.

How to read this

A result describes one version, in our setup, on one date. It is not a judgement of a vendor's intent. Where a vendor has replied to a finding, the reply is linked from the result. We do not accept payment from vendors. Vendors and users can ask us to include a harness by writing to research@agenticthinking.uk. Nobody pays.

Disclosure: Agentic Thinking is an independent research lab. We steward the open-source AgentHook standard and maintain the open-source HookBus project. Agentic Thinking also owns AgentProtect, commercial software available for licensing.