--- title: 'In the News: August 25, 2026' description: 'An evolution agent rewrote a coding harness and beat the hand-built Codex harness on Terminal-Bench 2: 69.7 to 77.0 percent in ten iterations.' canonical_url: 'https://darkfactory.dev/news/2026-08-25-evening' markdown_url: 'https://darkfactory.dev/news/2026-08-25-evening.md' collection: news date_published: '2026-08-25T19:28:00-04:00' date_modified: '2026-08-25T19:28:00-04:00' --- # In the News: August 25, 2026 An evolution agent rewrote a coding harness for ten rounds and raised Terminal-Bench 2 pass@1 from 69.7 to 77.0 percent, above the hand-built Codex harness at 71.9 percent. Each proposed edit carried a falsifiable prediction, and the system reverted changes that did not deliver the predicted result. ## 1. An evolve agent rewrote a coding agent's harness and beat the hand-built Codex harness **[Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses](https://arxiv.org/abs/2604.25850)** · Jiahang Lin, Shichun Liu, Chengjun Pan and colleagues; Fudan University, Peking University, and Shanghai Qiji Zhifeng · arXiv, April 28, 2026, revised May 18 AHE is a closed loop in which one agent edits the harness of another while the base model stays fixed. The harness is decomposed into seven component types held as plain files: system prompt, tool descriptions, tool implementations, middleware, skills, sub-agent configs, and long-term memory. Rollout trajectories are distilled into a layered evidence corpus, and every edit ships with a manifest naming the failure evidence it targets, the tasks it predicts it will fix, and the tasks it puts at risk. The next round grades that prediction, and edits that do not cash out are reverted at file granularity. Starting from a bash-only seed, with GPT-5.4 driving all three roles, ten iterations lifted pass@1, the single-attempt solve rate, on Terminal-Bench 2 from 69.7 percent to 77.0 percent, past the human-designed Codex harness at 71.9 percent and two self-evolving baselines, ACE and Training-Free GRPO. The frozen harness then transferred without further evolution: highest aggregate success on SWE-bench-verified at 12 percent fewer tokens than the seed, and gains of 5.1 to 10.1 points across three other model families. Ablations put the gain in tools, middleware, and long-term memory; the evolved system prompt inserted alone scored below the seed. The loop's weak side is regressions. Fix predictions landed roughly five times above a random baseline, regression predictions only about two, and the authors name "regression foresight" as the clearest direction for future loops while calling the system a controlled research prototype. [Code and the full prompts are public](https://github.com/china-qijizhifeng/agentic-harness-engineering). **Why it matters:** Practitioners still tune harnesses by hand. This paper offers an auditable way to delegate that work: require each edit to state a testable prediction and roll it back when the prediction fails. Its ablation also points to where tuning time paid off. The transferable value sat in tools, middleware, and memory rather than prompt prose. ## 2. Agents take their operating knowledge from instruction files, not API references **[From Agent Behaviour to Agent-Friendly Documentation](https://arxiv.org/abs/2608.20195)** · Zhijun Gao and Jing Chen, Peking University · arXiv, August 20, 2026 The study traces documentation behaviour across 557 real agentic coding sessions and 33,097 agent-authored pull requests, 94,813 development events in all. Instruction files and working notes account for 60.5 percent of the documentation interactions observed; API references, where most human documentation effort goes, account for 1.3 percent. The assumed read-the-docs-then-write-code sequence effectively does not occur, with an adjacent transition probability the paper puts at 0.002. The corpus contains no observed case of an agent explicitly validating its work against documentation, and documentation trails code changes 4.7 times more often than it leads them. **Why it matters:** If agents get their working knowledge from instruction files and notes, those files are the documentation surface worth engineering. In this corpus, documentation never served as a validation check, so teams must add that check deliberately rather than assume it already happens. ## 3. OpenAI deprecates the codex mcp-server command **[Codex MCP server command deprecated](https://developers.openai.com/codex/changelog)** · OpenAI · Codex changelog, August 24, 2026 The whole entry is three sentences: "The `codex mcp-server` command is now deprecated. Use the Codex app server instead. To use Codex from Claude Code, use the Codex plugin for Claude Code." No removal date is given. **Why it matters:** Anything that bridges Codex into another harness over MCP now has a migration on its calendar, and the sanctioned route for driving Codex from Claude Code is a vendor plugin rather than the open protocol surface. The missing removal date leaves the timing uncertain, but factory plumbing built on this command now carries migration work. --- ## Also this cycle - **[Trigger scheduled tasks from Gmail, Slack, and GitHub events](https://developers.openai.com/codex/changelog)** · OpenAI changelog, August 25, 2026 · ChatGPT scheduled tasks can now fire on GitHub pull-request activity, including reviews, comments, commit updates, and merges: a hosted trigger surface for agent runs on repository events.