$ cat briefs/daily/2026-10-15.md
2026-10-15
October 15, 2026 (Thu)
3 stories · 2 paper picks · 4 new pages
Top Stories
1. Long-running research needs an explicit continuation decision
OpenAI reports Third Moment Research running 10,000 agents concurrently on 36-48-hour investigations. An outer harness examines unfinished work and chooses retry, termination or human review. These are customer-reported deployments, not independent reliability measurements.
Why it matters: sustained execution depends on what happens when the inner agent stops. → Third Moment Research — research organization (source)
2. Amp adds voice coordination for other agents
Calling Puck uses GPT-Live-1 to let users launch and coordinate agents by voice, including from a watch. Amp is also testing Astra's Ultrafast mode; no latency measurement is published.
Why it matters: voice becomes a control surface for ongoing work beyond the coding session. → Amp — agent developer (source)
3. Anthropic finds misuse that disappears when judged one request at a time
Election-abuse testing found reliable single-request refusals but missed malicious campaigns split into benign-looking tasks. Anthropic says it added classifiers for multi-step misuse; no post-change success rate is given.
Why it matters: safeguards need to evaluate connected workflows as well as individual prompts. → AI Governance (source)
Paper Picks
LHTB 2.0: 77 tasks measure terminal-agent progress over three hours. Across 1,145 runs, GPT-6-astra leads at 0.69 dense mean reward. Read it for intermediate verification; dense reward is not a completion rate. → Long-Horizon-Terminal-Bench 2.0: measuring sustained progress (source)
ClawLoop: rollout diagnosis guides the next task-synthesis and SFT/RL cycle without an external teacher. Read it for closing the training-feedback loop; the abstract reports improvements without numerical scores or compute-matched comparisons. → ClawLoop: feedback closes the training cycle (source)
Watch
- Google and GovTech Singapore use a lightweight classifier followed by fine-tuned Gemma for on-device scam detection. Rollout will be phased and consent-based; accuracy and false-positive rates remain unpublished. → Google DeepMind (source)
New in Wiki
- Third Moment Research — research organization and Amp — agent developer — new organization pages for reader review. (source) (source)
- Long-Horizon-Terminal-Bench 2.0: measuring sustained progress and ClawLoop: feedback closes the training cycle — methods and evidence limits. (source) (source)
Updates
- Agent Runtime Containment records Daybreak testing before research simulations launch, with no measured residual escape rate. (source)