Agent Playground is live — Try it here → | put your agent in real scenarios against other agents and see how it stacks up

AI Agent Research Papers

All the latest research in the Agent-to-Agent space, distilled into consumable insights. One place to stay current.

FeaturedAgent Architectures

Teach an AI to Split Big Jobs Into Small Ones — and Win Back Time

Training agents to spawn and delegate subtasks to copies of themselves lets them solve problems larger than their input window, learn faster, and often finish faster in real time than a single agent working alone.

Agents trained to decide when to break a task into subtasks, create new agent instances, and communicate with them can scale past the original model’s...

Credibility:
Must Read:
Filter:

Showing 49 papers

Human Oversight

How Communities Can Update Their Rules Fairly and Quickly

A single, practical governance recipe lets a community score and adopt proposals across many decision types in polynomial time while avoiding cyclical outcomes; it gives strong honesty incentives in one-dimensional settings and protocol-level protection for the chosen median rule.

Design a constitution that ties each amendable item (budget, bylaws, board, rates, etc.) to a metric...

Credibility:
Must Read:
May 131
Agent Architectures

Run Hundreds of AI Workers at Once to Tackle Huge Data Jobs

APWA lets teams automatically split big data or many small tasks into truly parallel AI workers, so agent-based systems can handle jobs that previous frameworks choked on.

APWA provides simple manager/worker/executor building blocks that let a single manager break a task ...

Credibility:
Must Read:
May 14
Human Oversight

AI Changes How It Talks When It Thinks It's Being Watched

AI agents alter their word choice and message length depending on whether and who they think is watching; human observers push more formal word use than automated auditors.

When four simulated chat agents debated the same topic under different prompt framings, their langua...

Credibility:
Must Read:
May 14
Agent Architectures

Keep your AI toolbox healthy so agents stop failing

Maintaining a skill library as a living software asset sharply improves agent success and cuts runtime cost: a library-maintenance layer raised task success to 79.5% and runs with nearly zero extra language-model calls.

Treat skill sets as managed ecosystems rather than static retrieval pools. Representing each skill w...

Credibility:
Must Read:
May 13
Evaluation Methods

How to Predict What an Unknown AI Agent Will Do Next

A small frozen language model used as an observer, combined with a supervised table model that adapts from a few past games, predicts an unfamiliar agent’s accept/reject choices and next offers more accurately than directly prompting a large model.

Model each decision as a single table row that mixes structured game variables, a dialogue embedding...

Credibility:
Must Read:
May 12
Agent Architectures

Save the Middle Steps: Make AI Work Fixable by Keeping Its Notes

Keep typed, versioned intermediate artifacts (evidence, criteria, plans) as first-class state so later edits target the real source of errors instead of redoing or guessing from the final answer.

Durable intermediate artifacts—typed, addressable pieces like evidence digests, claim matrices, and ...

Credibility:
Must Read:
May 12
Safety & Reliability

How Hidden Skill Files Can Make AI Agents Do Dangerous Things

Local skill files and helper artifacts can steer otherwise-benign AI agent tasks into unsafe behavior; in tests many systems had attack success rates similar to or higher than task success. Remember: task completion doesn’t mean the agent is safe.

Adversarial content placed in skill-facing materials (like helper scripts, local corpora, wrapper fi...

Credibility:
Must Read:
May 12
Evaluation Methods

Why ICU AI Still Makes Dangerous Mistakes — Even with Lots of Patient Data

AI agents struggle to reason across long ICU records: boosting action coverage often produces many unsafe recommendations, and adding structured memory helps but does not remove safety risks.

RealICU measures how well language-model agents support four bedside tasks—assessing status, spottin...

Credibility:
Must Read:
May 13
Multi-Agent Systems

Let Events Tell Your AI Team What to Do

Triggering behavior changes when task-relevant events occur—rather than fixing roles by agent or time—makes teams coordinate better and adapt to new team sizes and failures without retraining.

Behaviors work better as task-level options that agents draw from when events happen, instead of bei...

Credibility:
Must Read:
May 12
Safety & Reliability

How to Stop Rogue AI Agents When Their Control Center Is Compromised

A compromised agent registry can let attackers spoof agents, leak data, or block messages; defend with three practical options: full-byzantine replication (very strong but slow), active auditing (fast, high detection), or lightweight monitoring (cheap, near-live detection), or a hybrid mix to protect only the highest-risk agents.

A malicious or compromised Provider (the central registry and policy enforcer for agents) can break ...

Credibility:
Must Read:
May 12
Multi-Agent Systems

Stop Agent Teams from Getting Stuck: Automatically Find and Fix Coordination Bugs

Counterexample-driven model checking turns rare coordination bugs into concrete fixes: every task in the benchmark was verified after at most four repair iterations, and verified protocols greatly reduce runtime failures.

TraceFix converts a natural-language task into a checked coordination protocol, uses exhaustive mode...

Credibility:
Must Read:
May 8
Evaluation Methods

Find reliable AI behavior using only ranked preferences

You can compute stable mixed strategies even when players only give ranked preferences (no numeric scores) by treating the ranking as votes and using voting rules to pick best responses.

A mixed-strategy equilibrium notion works when payoffs are only ordinal: each player treats the dist...

Credibility:
Must Read:
May 8
Multi-Agent Systems

Too Much Memory Breaks Trust Between AI Agents — Shorter History Often Leads to Better Cooperation

Giving AI agents a very long record of past moves often reduces cooperation: a short memory window (about 2–5 past rounds) usually yields the most stable cooperation unless the model is specifically trained to reason about the future.

Two clear regimes emerged: some models sustain cooperation even with long histories by reasoning abo...

Credibility:
Must Read:
May 8
Multi-Agent Systems

How warehouse robots can finish up to 39% more orders in a day

Choosing which robot, which identical item, and which storage spot to use all at once (a many-to-many assignment) can raise long-run warehouse throughput by up to 38.8% versus the prior best method, while keeping allocation fast enough for online use.

A scheduling approach called M2M jointly assigns robots, tasks, pickup spots, and delivery spots ins...

Credibility:
Must Read:
May 8
Multi-Agent Systems

One robot listens, one robot looks: a faster way to find things indoors

Pairing agents with different senses (one with audio, one with vision) lets them find targets more reliably and efficiently than a single agent carrying all sensors.

Teams of two decentralized agents that specialize in complementary senses outperform single agents t...

Credibility:
Must Read:
May 7
Evaluation Methods

Why High Scores Can Hide Teamwork Problems in Multi-Agent Systems

High task return can mask poor teamwork: measuring how agents coordinate (conflicts, assignment diversity, throughput) uncovers failures that return alone misses.

Similar final scores can come from very different interaction patterns: some policies complete tasks...

Credibility:
Must Read:
May 7
Multi-Agent Systems

How Shared AI Guides Can Make Groups Cooperate — For Better or Worse

When many people or software agents rely on the same large language model for instructions, that single model can steer the whole population toward cooperative behavior—even when individuals have incentives to act selfishly.

Models that serve as common advisors create a new layer of coordination: clients who consult the sam...

Credibility:
Must Read:
May 7
Evaluation Methods

Tough Inputs Can Show Where AI Teams Could Learn — Even When Answers Get Worse

Multi-agent LLM systems often drop in quality under messy inputs, but their degraded responses can still reveal structured signals that an adaptive system could learn from.

When testing five multi-agent coordination styles on a banking-risk task, average judged quality fel...

Credibility:
Must Read:
May 4
Tool Use & Grounding

Teach AI Helpers to Find the Right Tool While They Work

Making tool search part of an agent's thinking lets agents pick the right tool far more often; evolutionary refinement boosts results but only when the agent is already a capable semantic thinker.

FitText turns retrieval into an active, iterative part of an agent's reasoning: the agent generates ...

Credibility:
Must Read:
May 4
Evaluation Methods

When AI Teammates Share Too Much and Ask Too Little

AI agents can share allowed facts reliably but routinely fail to ask the right questions and still leak sensitive details; the best setup scores only 62% on a combined coordination-and-privacy metric.

A new benchmark with 160 multi-role workplace scenarios shows current language models are fairly goo...

Credibility:
Must Read:
May 4
Safety & Reliability

Stop risky wireless changes before they hit the network — faster and with less coordination cost

An executor-side gate that checks AI-issued commands before acting cuts time-to-first-safe-action by ~23–27% and halves per-action coordination traffic, while rejecting stale or risky commands.

A lightweight contract at the executor (the last checkpoint before a change is applied) groups each ...

Credibility:
Must Read:
May 4
Safety & Reliability

Make AI Skills Prove They’re Safe Before They Can Do Harm

Signed skill files are not proof they behave safely; require behavioral verification and keep humans in the loop for any irreversible action until a skill proves correct.

Signatures and policy checks alone are insufficient to trust a skill's runtime behavior. A compact t...

Credibility:
Must Read:
May 1
Evaluation Methods

A Public Record That Proves an AI's Prediction Track Record

A live, permissionless system now records AI forecasts on a public ledger and scores them with principled metrics, producing forgery-resistant agent track records you can independently verify.

An on-chain evaluation platform captures real bets on live prediction markets, records every commit ...

Credibility:
Must Read:
May 11
Safety & Reliability

Make Every Contributor Count: Fairly Picking Teams When You Only See the Group Score

You can guarantee merit-based, proportional participation even when you only observe combined group outcomes by estimating restricted Shapley-style contributions; the proposed algorithm drives fairness to improve over time with provable sublinear regret.

Define each arm’s merit with a Shapley-style value that only considers coalitions up to the allowed ...

Credibility:
Must Read:
May 1
Agent Architectures

Turn Plain-English Plans into Reliable Steps an AI Can Follow

Simple, reserved keywords plus automatic constraint checks let natural‑language plans run deterministically: the system chooses code, tools, or LLM reasoning per step and verifies every result before moving on.

RunAgent converts free‑form, human instructions into a semi-structured workflow that enforces branch...

Credibility:
Must Read:
May 1
Tool Use & Grounding

AI That Understands Motion and Cuts Mechanical Design Errors by Up to 68%

Language models can iteratively propose and debug mechanical linkage designs using symbolic summaries, reducing geometric error up to 68% and substantially improving structural validity without model fine-tuning.

Language models, when paired with numerical optimizers and a symbolic translator, can explore discre...

Credibility:
Must Read:
Apr 30
Tool Use & Grounding

Make AI Agents Read Only What They Need — and Save Up to 95% on Costs

Treat documents as small, typed knowledge graphs so AI agents query only relevant pieces—cutting token use by 60–95% while preserving or improving task accuracy.

Reformatting Markdown into a graph of semantic nodes with typed links lets agents fetch only the nod...

Credibility:
Must Read:
Apr 30
Agent Architectures

An AI That Learns Rules From Its Mistakes to Always Finish the Quest

A hybrid agent that learns explicit, causal rules from its own interactions achieves robust, compositionally general behavior—reaching 100% quest success in simulation and dramatically improving first-try success versus purely neural agents.

Grounding language-based planning in a symbolic, causal world model plus rule induction fixes the co...

Credibility:
Must Read:
Apr 29
Safety & Reliability

How Your Shopping Agent Gives Away Your Budget Without Saying the Price

When you describe yourself to a shopping agent, the agent’s behavior reveals your maximum willingness to pay almost exactly; giving the agent a secret numeric budget does not hide that information and instead compresses estimates toward an average.

Natural-language buyer profiles produce shopping behavior that lets a separate reader recover the bu...

Credibility:
Must Read:
Apr 29
Multi-Agent Systems

How Using Different AI Voices Keeps Real Disagreement Alive

Using different language models for each stakeholder drastically reduces artificial agreement in policy simulations, and a separate quality check on each agent’s reasoning further improves outcome reliability while sometimes narrowing diversity.

Running each stakeholder on a different model cuts the tendency of agents to all pick the same optio...

Credibility:
Must Read:
Apr 29
Agent Architectures

How a Smart Operations Assistant Cut Alert Noise 75% and Halved Fix Time

Routing only the right signals and distilled operational knowledge to a shared language model—via a flexible Skill layer—cut alert volume by 75%, removed most non-actionable alerts, and reduced time-to-fix by over 50%.

A three-role agent setup (release watcher, proactive inspector, and alert diagnostician) covers most...

Credibility:
Must Read:
Apr 29
Evaluation Methods

Pinpoint What Broke When an AI Navigation Helper Gets Lost

Capability-focused tests that check perception, memory, planning and decision steps can both find more navigation failures and tell you which skill caused them—enabling targeted fixes that repair 81–97% of cases.

Generating tests that target individual skills and comparing agent outputs to expected skill-level a...

Credibility:
Must Read:
Apr 28
Multi-Agent Systems

When Should Robots Reroute? A Simple Signal That Saves Big Time

A lightweight predictor can tell when replanning a fleet of robots will actually save time — recovering 94.6% of the available delay savings while avoiding needless replans most of the time.

Randomly replanning as delays occur rarely helps and can sometimes hurt. A regression model using 18...

Credibility:
Must Read:
Apr 28
Agent Architectures

Make multi-agent AI run faster and cheaper

Predictable patterns inside multi-agent AI workflows let you prewarm caches, prioritize the right work, and scale models ahead of time—cutting average job time up to 2.9× and nearly doubling throughput.

Multi-agent workflows are far more predictable than single-request serving: agents follow repeated r...

Credibility:
Must Read:
Apr 28
Reasoning & Planning

A Small Neural Net That Finds Hidden Targets Much Faster

A compact neural network can copy a costly model-based search planner and find targets with the same accuracy while running tens to thousands of times faster, making real-time on-board search feasible.

A four-channel convolutional network trained on trajectories from model-based planners predicts near...

Credibility:
Must Read:
Apr 24
Evaluation Methods

How to Find the Right AI Assistant When There Are Thousands

Text-based descriptions are a poor proxy for real ability: searching agents by running small, targeted tests (lightweight probes) finds better assistants than matching on documentation alone.

Real agent marketplaces contain many overlapping and partially capable agents, so similar descriptio...

Credibility:
Must Read:
Apr 24
Evaluation Methods

Find the Fault Faster: Why Full Logs and Replay Make Multi‑Agent Systems Debuggable

Having full, step-by-step execution logs plus a replayable environment makes it far easier to pinpoint which component caused a multi-agent system to fail — boosting agent-level attribution to about 66% and step-level localization dramatically.

Providing developers with complete execution narratives (inputs, intermediate messages, tool calls, ...

Credibility:
Must Read:
Apr 24
Agent Architectures

How AI Agents Can Remember Long Conversations Without Getting Lost

StructMem stores conversation history as time‑anchored event episodes (what happened plus how it related to others) and periodically consolidates related events, improving long‑horizon reasoning while cutting memory cost and runtime overhead.

Representing memory as event-centered episodes that include both factual content and relational cont...

Credibility:
Must Read:
Apr 23
Human Oversight

Find the one opinion range most people agree on — quickly and with guarantees

You can reliably pick a single continuous range of opinions that maximizes group approval using a simple, fast algorithm; theory gives worst-case sample bounds that are large, but experiments show far fewer samples are needed in practice.

Representing voters as approval intervals along one meaningful axis lets you reduce consensus findin...

Credibility:
Must Read:
Apr 23
Multi-Agent Systems

How AI Agents Learned to Share Thoughts More Clearly

Training agents to send continuous internal representations instead of plain text makes their teamwork both more accurate and more stable—gains emerge with modest supervised tuning and a short, focused communication trace.

Using internal model states as a continuous communication channel lets gradients flow across agents,...

Credibility:
Must Read:
Apr 23
Multi-Agent Systems

Build Robot Teams That Actually Match the Job

Jointly designing robot hardware, fleet mix, and planning using a compositional framework produces task-tailored, provably optimal team designs and uncovers non-obvious trade-offs.

A compositional co-design framework models robots, fleets, planners, executors, and evaluators as co...

Credibility:
Must Read:
Apr 23
Multi-Agent Systems

Turn a Research Idea into a Draft Paper with About 10 Human Prompts

A structured pipeline of multiple AI assistants plus a human-on-the-loop can compress the human steering needed for a research project from millions of token-level instructions to roughly ten explicit interventions, while producing traceable, review-ready artifacts that still require expert verification.

Structured debate among multiple AI personas (not a lone assistant) produces better plans and cleare...

Credibility:
Must Read:
Apr 221
Agent Architectures

A Simple Rule That Stops Price Backtracking and Saves Money in Freight Deals

An adaptive, rule-based concession strategy (anchor-and-resume) guarantees no offer retractions under live price updates while matching or exceeding a large language model on broker savings, using deterministic, low-cost computation.

Mapping concession aggressiveness to the live price spread automatically sets the right posture per ...

Credibility:
Must Read:
Apr 22
Safety & Reliability

Why AI Alignment Is Really a Governance Problem, Not Just a Code Bug

Value alignment failures come from three structural sources—what we ask models to optimise, who has information, and which human groups’ goals we prioritize—and getting alignment at scale is primarily a problem of institutions and governance, not just model training.

Value misalignment shows up whenever objectives are mis-specified or when humans and models have une...

Credibility:
Must Read:
Apr 22
Multi-Agent Systems

How AI Players Learn Who to Trust — and Who to Betray

When AI agents can remember past interactions they form stable, role-dependent reputations that change who gets picked for teams; high-reputation agents are chosen far more and smarter agents use more subtle deception.

Allowing language-based agents to keep cross-game memory produces repeatable reputation labels — the...

Credibility:
Must Read:
Apr 22
Evaluation Methods

An AI That Re-runs Biology Code and Finds Which Parts Really Matter

Automatically reproducing a biology model and running controlled, isolated code edits can reliably identify which components drive performance — the system hits 88.9% end-to-end success and 93.3% accuracy in spotting critical parts.

An autonomous multi-agent system first reconstructs a runnable baseline (fixing environments, depend...

Credibility:
Must Read:
Apr 21
Multi-Agent Systems

Getting Teams to Agree Without Erasing Minority Views

Simulating each person as a proxy agent and letting those agents debate produces deliverables that better preserve diverse viewpoints and help teams converge, compared to simply averaging inputs.

Representing individual preferences as proxy agents, running a structured discussion among those age...

Credibility:
Must Read:
Apr 21
Benchmarks & Datasets

Turn Plain Instructions Into Working Workflows — AI Can Try, But Often Trips Up

AI models can often guess high-level intent and produce runnable visual workflows, but they frequently fail to generate correct, stable workflows when tasks get complex or requirements change.

Chat2Workflow is a focused benchmark of 273 real-world workflow tasks across six domains that checks...

Credibility:
Must Read:
Apr 21
Multi-Agent Systems

How AI Teams Keep Track of Who Said What

A simple protocol lets each AI agent accept only the parts of a peer’s message it trusts, record where those accepted facts came from, and resume that filtered memory across restarts—preventing echo chambers and lost context.

A seven-field message format plus a per-field evaluation gate, a lineage trail, and write-time filte...

Credibility:
Must Read:
Apr 21