The claim survived, the caveat did not

Three papers on what survives the handoffs inside an agent system: the authorization caveat a summarizer quietly drops, the subtasks a second agent should never have been handed, and the benchmark runs you can stop paying for.

Today's pick

9% → 98%

approval of risky actions by one monitor once the word unverified was dropped from an otherwise identical handoff

Silence Is Endorsement: Verification-Status Laundering in LLM Agent Pipelines

The safety check in most agent systems reads a summary, not the original evidence — so when a planner's note that an action is authorized, unverified passes through a summarizer and a memory compressor, the claim survives and the caveat does not. Holding the action and the authorization wording fixed and removing only the unverified framing raised one monitor's approval of risky actions from 5% to 60% and another's from 9% to 98%, and a full proposer-summarizer-memory-monitor chain reached 57% to 81% risky approval across three downstream monitors; memory compression alone dropped the marker in 60% to 100% of items depending on which model compressed. Instructing the monitor to reject unverified authorization is not a reliable patch — two models stopped being fooled, one got worse, and one began refusing legitimate requests. Carry authorization as structured state attached to the specific claim and have the gate read that field, instead of trusting a caveat that later stages are free to paraphrase away.

open the explainer → abstract PDF

71.8 vs 53.6

score when subtasks barely interact, splitting the work across agents versus keeping one — and 75.5 vs 76.7 when each step feeds the next

Rethinking Multi-Agent Collaboration: When More Is Less

The reflex when an agent task gets hard is to add more agents, and this paper draws the boundary: splitting work pays off on long tasks whose pieces barely interact, and costs you on tasks where each step consumes the last one's output, because the split cuts a chain the single agent was holding in one context. Their scheduler spawns a sub-agent only at a real seam — a dependency whose removal would actually separate the work into independent halves — and lets branches pull each other's results by retrieval instead of broadcasting context, scoring 71.8 against a single agent's 53.6 on a day-long instruction-following task while staying level (75.5 against 76.7) on terminal work where steps chain. Scaling past that hurt on the same benchmarks: four agents to six dropped the score from 71.8 to 64.0 and nearly doubled input tokens, and letting sub-agents spawn their own sub-agents lost points at every extra level. Before adding a second agent, look for the seam; if the work has none, you are paying coordination overhead for a worse result.

open the explainer → abstract PDF

22 of 113

benchmark tasks whose single run tracked the full benchmark score reliably enough to decide on

DeltaSelect: Affordable A/B Testing for Coding Agents

Running a full coding-agent benchmark every time you change a prompt or a tool is too slow to iterate on, but a handful of tasks picked by eye is noise. Resampling one run per task from published trials showed only 22 of 113 tasks tracked full-benchmark performance reliably enough to read from a single run, so the method ranks tasks by that conservative correlation, converts each task's fraction of tests fixed onto the benchmark's scale by regression, and fills a fixed dollar budget down the ranking. Across 13 evaluations on the eight tasks that survived, costing $27.86 in total, it carried an agent-instruction rewrite that cut cost per run by 58.1% — solid at the significance level used — while the score moved from 36.5% to 42.4%, which was not. Two warnings worth stealing: the published baseline for those tasks was 21 points off once they ran in the author's own harness rather than the benchmark's, so re-baseline in the harness you actually ship; and freeze the task set before you start comparing, or each week's number is a different benchmark.

open the explainer → abstract PDF