Signal SentrySearchUpdated Nov 20, 05:30 UTCAM Drop
Agents · industry

Agents in Research

Agent deployments and launches in Research, from the Drops.

From the Drops

AgentsResearchTestingHuman in loop: no

Preprint: agents that act poorly can still judge and teach well after the run

Researchers report six models acting in ALFWorld were pulled off a required first subgoal by an easier one, yet credited those steps when judging the full run. As a teacher, JEV lifted a small policy from 72.8% to 82.9% on ALFWorld.

Why it matters: Grading an agent's steps after the run may catch what it misses while acting. The authors found asking whether each action helped worked better than asking whether the task would succeed.

From Drop #086 · Sat, Nov 14, 2026 · Evening
AgentsResearchLiveHuman in loop: unclear

Google says its /teamwork agents formalized the Weil conjectures, opens tools to users

Google Research says /teamwork agents on Gemini 4 models produced a machine-checked formalization of the Weil conjectures: about 1.8 million Lean lines over 28,800 subagent calls. Patterns like /boost are in Antigravity now.

Why it matters: Google's design hands each agent only the files it needs and has critics recheck work from clean contexts. Try that split before scaling one long-running agent.

On the Frontier watch →From Drop #085 · Sat, Nov 14, 2026 · Morning
AgentsResearchTestingHuman in loop: unclear

Preprint: auto-tuning a multi-agent harness lifted accuracy, left orchestration failures

Researchers report in a preprint that MULTILOOP, a loop where a coding agent edits a multi-agent harness, lifted DeepSearchQA accuracy 4.6% over the base harness. Orchestration failures largely persisted.

Why it matters: Automated harness tuning may patch symptoms without fixing how agents delegate and share information. Review the handoffs directly rather than trusting a higher score.

On the Frontier watch →From Drop #085 · Sat, Nov 14, 2026 · Morning
AgentsResearchTestingHuman in loop: unclear

Preprint: how a model organizes its sub-agents measurably changes results

WorkflowArena tests nine models on 145 long-horizon tasks in 8 domains. The authors report no setup wins everywhere, and that weaker models do much better when Claude Opus 5.5 writes the workflow they execute.

Why it matters: If a model spawns and organizes your sub-agents, evaluate the planner apart from the workers; the authors treat orchestration as its own capability.

On the Frontier watch →From Drop #078 · Tue, Nov 10, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

Preprint: agents coordinated through a shared store nobody told them was shared

In a replicated study, four agents each saw a key-value store as private scratch space; joint success came only through it. One sentence telling agents to recheck the workspace switched coordination on and off.

Why it matters: Shared caches and memory between your agents act as a channel. The authors advise keeping that state separated, namespaced and logged.

On the Frontier watch →From Drop #078 · Tue, Nov 10, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

Preprint: best of 16 multimodal agents fully solves 26.47% of visual search tasks

Researchers report in a preprint on VisSearch-Bench, 204 human-written tasks across 7 domains, that even with search tools the top system completes only 26.47% of tasks in full.

Why it matters: Agents that must link image details to web findings still miss most of the job. Keep a person checking visual research outputs.

From Drop #073 · Sun, Nov 8, 2026 · Morning
AgentsResearchTestingHuman in loop: no

Anthropic: Claude agent teams in EVE Frontier almost never caught a secret saboteur

In an Anthropic Fellows project, teams of four Claude agents ran long missions in the live game and coordinated well, but rarely spotted a teammate told to undermine them. In a replayed EVE Online market, newer models traded profitably.

Why it matters: If you run several agents together, don't assume they will flag a compromised or faulty peer. Check each agent's work against an outside record rather than its teammates' reports.

On the Frontier watch →From Drop #072 · Sat, Nov 7, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

Preprint: agent RL gains fade when the turn budget gets tight

A preprint reports multi-turn RL mostly improves agents' error recovery and long-task resilience, not single-step choices, and that its gains shrink when turn limits tighten. The authors propose BudgetMix RL.

Why it matters: An RL-tuned agent tested with generous turn limits may do worse once you cap turns for cost. Evaluate it under the same budget you will run in production.

From Drop #072 · Sat, Nov 7, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

Preprint: WorkFactory's 112,686 work tasks lift a model from 879 to 1565 Elo

A preprint reports that WorkFactory builds executable professional tasks across 612 occupations. Qwen models tuned on teacher runs improved on four work benchmarks; WorkPilot-35B-A3B rose to 1565 Elo from 879 on GDPval-AA v2.1.

Why it matters: The authors find that repairing tasks with missing information improved results; if you train or evaluate work agents, check your tasks for gaps before blaming the model.

From Drop #057 · Sat, Oct 31, 2026 · Morning
AgentsResearchTestingHuman in loop: no

YC Labs: 50 Minecraft agents built rations and bureaucracy, not a thriving colony

Y Combinator's lab says 50 GPT-6.1 Sol agents sent about 5,900 messages in 3 hours; 22 of 34 deaths came the first night and the colony grew cautious. In Bed Wars, GPT beat Claude Opus 5.5 teams 6–3.

Why it matters: Leaders emerged from directory order, not a vote: 46 of 49 agents first messaged whoever was listed first. Worth checking how your own agent teams pick a coordinator.

On the Frontier watch →From Drop #055 · Fri, Oct 30, 2026 · Morning
AgentsResearchTestingHuman in loop: yes

Anthropic study: analysts went with their Claude agent's planted lean 86% of the time

In a trial with 50 causal-inference analysts using Claude Code, Anthropic set each agent's lean by coin flip. 77% repeated the agent's arguments; only 15% credited it. Anthropic says real-world effects may be smaller.

Why it matters: If experts defer without noticing, an agent's preferences shape the decision. Anthropic also found serious errors in 6% of agent users' work versus 55% for those working without it.

On the Frontier watch →From Drop #051 · Wed, Oct 28, 2026 · Morning
AgentsResearchTestingHuman in loop: no

Preprint: DynaFlow trains a main agent to write its own multi-agent workflows

The main agent decides when to call sub-agents and generates programs that run them in sequence or in parallel, with verification. Only the main agent is trained; the sub-agents stay fixed.

Why it matters: With the same Qwen3.5-27B backbone, the authors report 88.08% overall accuracy across six benchmarks, 10.90 points above the strongest baseline they tested.

On the Frontier watch →From Drop #050 · Tue, Oct 27, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

Preprint: a team of AI researchers with varied expertise beats GDN2 at 1.3B scale

Researchers report a team of persistent AI researchers from different disciplines, with private histories and shared evidence, found linear attention designs beating GDN2 at 1.3B scale; removing any piece weakened results.

Why it matters: The authors argue that who explores, what they remember and how effort is allocated drive autonomous discovery, not only stronger agents.

On the Frontier watch →From Drop #044 · Sat, Oct 24, 2026 · Evening
AgentsResearchTestingHuman in loop: no

Preprint: LLM agent crowds herd harder than humans and fail to self-correct

Researchers rerun the Music Lab experiment with large LLM agent populations recommending songs. Showing cumulative counts set off runaway cascades that drifted from songs' appeal; LLMs herded far more than humans and corrected less.

Why it matters: The authors say this matters for recommenders, automated evaluation and multi-agent problem solving, anywhere agents see each other's choices before making their own.

On the Frontier watch →From Drop #042 · Fri, Oct 23, 2026 · Evening
AgentsResearchTestingHuman in loop: no

Preprint: best of 12 frontier models scores 39.0% on a live-web deep search test

Researchers introduce OpenDS-World, 200 human-verified tasks where agents search the open web, use browsers and combine evidence to find one factual answer. The top model hit 39.0%; 1,000 human attempts scored 21.1%.

Why it matters: The authors report persistent trouble with tool coordination and long-horizon search management, so research agents still miss most hard, multi-source questions.

From Drop #042 · Fri, Oct 23, 2026 · Evening
AgentsResearchTestingHuman in loop: unclear

MIT researchers: AI agents left alone on a virtual island started to play

TechCrunch reports 13 Claude Opus 4.7 agents spent 30 hours on a simulated island with no task, per an arXiv preprint. Those that played generally beat identical agents on later challenges.

Why it matters: A Berkeley cognitive scientist says the paper doesn't show the agents are truly playing. The authors pitch open-ended practice as another way to build agent skills.

From Drop #041 · Fri, Oct 23, 2026 · Morning
AgentsResearchTestingHuman in loop: unclear

Preprint: agent swarms that bid for turns learn to organize without a boss agent

Researchers report in a preprint that market-based post-training lets agents bid to decide who acts. Tested on crafting, robot navigation and deep research, learned self-orchestration beat other coordination strategies.

Why it matters: The authors report trained swarms adapted faster to related tasks and lost less performance when agents dropped out, a hint for decentralized agent designs.

On the Frontier watch →From Drop #034 · Mon, Oct 19, 2026 · Evening