Self-Harness, from Hangfan Zhang and colleagues at the Shanghai AI Laboratory, has the agent mine model-specific failure patterns from its own execution traces, propose minimal harness edits tied to those failures, and keep only the edits that survive regression testing (arXiv, submitted June 8). Starting from a minimal harness, three base models (MiniMax M2.5, Qwen3.5-35B-A3B and GLM-5) improved on every held-in and held-out split of Terminal Bench 2.0, SWE-bench Verified and AppWorld, with relative gains of up to 132% and no human or stronger external agent in the loop. The newsletter leads with it as the Singularity “editing its own source code”, and it is the item that gave the harness writes itself its name. It follows Poetiq’s meta-system building its own harnesses and Meta’s hyperagents, and is followed the next day by Codex generating 99.8% of its own output tokens and in July by Kimi K3 rewriting its own Cline harness for eleven Terminal Bench points.
Development
An agent mines its weaknesses and rewrites its own scaffolding
The newsletter's "editing its own source code" moment: the scaffolding layer, until now the last reliably human-authored part of an agent, is rewritten by the agent from its own failure traces, which is what the self-authored-scaffolding theme is named for.
Claim
A new Self-Harness paradigm lets an agent mine its own weaknesses and rewrite its scaffolding, lifting Terminal-Bench scores by double digits with no human engineers involved.
Reported metric: double digits on Terminal-Bench
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to June 25, 2026 (theinnermostloop.substack.com)
- Self-Harness: Harnesses That Improve Themselves (arxiv.org)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- A Chinese lab says its model participates in its own evolution
- Hyperagents fuse task-solving and self-modification into one program
- Users metaprompt an agent to write its own objective file
- A coding agent generates 99.8% of its own weekly output tokens
- Agents act as their own data scientists, gaining most by optimizing themselves
- A skill ships for meta-harnesses that rewrite their own scaffolding
- An outer-loop agent rewrites its inner researcher seven times unattended
- A self-optimizing optimizer claims state of the art with no human intervention
- Welcome to June 25, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.