self.md radar — 2026-05-09
The useful part of today’s agent stack was bookkeeping: how long a run can hold, who wrote which line, and when a permission boundary hid evidence.
METR put hours on model-agent work. re_gent put blame on tool calls. Partial Evidence Bench put a name on a very corporate failure mode: the answer is allowed, the missing evidence is not, and the system acts as if nothing is missing.
1. METR stretched the agent stopwatch
sources:
what happened: METR updated its task-completion time horizon page on May 8 with Time Horizon 1.1, measured across more than 100 software tasks. The page defines the 50% horizon as the human-expert task length where an agent is predicted to succeed half the time; early Claude Mythos Preview is listed at 1,044.8 minutes for p50 and 185.9 minutes for p80. METR also leaves the small poison pill in the chart: measurements above 16 hours are unreliable with the current task suite.
why this matters: A benchmark like this does not say an agent can own your week. It says the honest unit is no longer “can it solve the task” but “how long before its reliability falls off a cliff, and who notices before the cliff gets expensive.”
2. re_gent adds blame to agent work
sources:
what happened:
re_gent calls itself “Git for AI Agents,” which is the least annoying version of that pitch because the README shows the actual ledger. After rgt init, the agent can run normally, while re_gent captures tool calls and exposes rgt log plus rgt blame src/file.go:42. The blame example ties a line to a step id, a Claude session, the tool used, and the prompt that produced it.
why this matters: Normal git answers what changed. Agent git has to answer whose loop changed it, under what instruction, and where to rewind when the diff was technically green but operationally cursed.
3. scoped evidence got a benchmark
sources:
what happened: Partial Evidence Bench targets a failure mode that will be everywhere in enterprise agents: access control works, but the final answer pretends the hidden documents were not part of the question. The benchmark ships 72 tasks across due diligence, compliance audit, and security incident response, with ACL-partitioned corpora and oracle complete answers, authorized-view answers, completeness judgments, and gap reports. Its surfaces are not fancy: correctness, awareness of missing evidence, gap-report quality, and unsafe completeness behavior.
why this matters: The nasty version of a permission boundary is not a denial. It is a clean answer that should have said, “i can only see half the file.” Silent filtering is how a compliant system lies without breaking policy.
supporting links
- Chrome DevTools MCP — browser agents get traces, network requests, screenshots, console messages, memory snapshots, and a README warning that MCP clients can inspect or modify browser data.
- From History to State — recurring workflows move from long prompts into lightweight modules; Qwen3-8B SFT+RL reports 89.6% unseen ALFWorld, 76.8% WebShop, 66.4% SciWorld, and 2-7x fewer prompt tokens.
- Armin on local models — the complaint is boring and correct: local coding models need provider-grade setup and tool-parameter streaming, not another quant victory lap.
- CodeGraph — claims 92% fewer Claude Code tool calls and 71% faster exploration across six codebases by pre-indexing symbols and call graphs.
left on the table
- METR’s Mythos measurement was the allowed repeat because it is a fresh benchmark/data drop; Mozilla’s Firefox hardening post stayed out because yesterday already covered that security-agent surface.
- DeepSeek-TUI and agent-skills were exact ledger repeats from May 8.
- Chrome DevTools MCP could have carried a fourth signal, but it worked better as support for the custody/debugging theme.
- The Obsidian music library map was a good wildcard, just not strong enough to beat measured failure surfaces.