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.

FeaturedMulti-Agent Systems

AI Teams That Learn to Run Scientific Analyses — 43% Success on a Tough Benchmark

Evolving how multiple AI agents are organized and scored meaningfully improves success on computational science tasks: Mimosa reached a 43% task success rate by iteratively mutating and evaluating workflows, producing stable, auditable pipelines after several rounds.

Iteratively changing and testing multi-agent workflows lets an automated system discover better ways to complete data-driven scientific tasks. Using a...

Credibility:
Must Read:
Filter:

Showing 49 papers

Multi-Agent Systems

Reassigning Tasks Cuts Robot Fleet Costs in 89% of Tests

Flexible task reassignment with localized path repairs finds better team plans than fixing which robot does each job — improving 89.1% of benchmark instances and yielding up to a median 12.2% cut in travel cost for the best variant.

A two-level search that repeatedly reassigns groups of related tasks and repairs the affected paths ...

Credibility:
Must Read:
Mar 30
Multi-Agent Systems

When Many AI Researchers Work Together: Faster Fixes or Deeper Breakthroughs?

Parallel pools of simple agents find quick, reliable improvements at scale, while small coordinated expert teams produce fewer but more diverse and complex code changes—combine both for best results.

Parallel "subagents" that run many short experiments in isolation deliver rapid, stable improvements...

Credibility:
Must Read:
Mar 31
Multi-Agent Systems

A Safe Real-Time Sandbox to Watch Social Media Bots in Action

You can reproduce realistic social media dynamics in a contained environment: agents grounded on live feeds behave like real users, letting researchers study misinformation spread without risking real people.

BotVerse runs simulated social accounts that react to live social-media trends while staying inside ...

Credibility:
Must Read:
Mar 31
Agent Architectures

Agents That Remember: How AI Learns Over Time and Gives New Agents a Head Start

Agents that persistently record and reuse interaction experience get substantially better over time, and that stored experience can be handed to a fresh agent to give it an immediate performance boost.

A runtime that treats collaboration, identity, and lifelong learning as first-class features lets ag...

Credibility:
Must Read:
Mar 30
Observability

See What Your AI Desktop Helper Actually Did — and Fix It

Users need clear, post-task traces of what AI helpers did — not just warnings — so they can spot risky actions, understand lasting changes, and undo or clean up effects.

Many people sense that AI desktop helpers are risky but cannot name what those risks are or what the...

Credibility:
Must Read:
Mar 30
Multi-Agent Systems

Train Teams of Agents from Saved Data Without Breaking Their Coordination

Adaptive, state-dependent partial action replacement lets you learn coordinated multi-agent policies from fixed datasets while cutting computation and reducing unreliable value estimates.

Learning how many agents to re-sample at each update (rather than trying every option) gives most of...

Credibility:
Must Read:
Mar 30
Evaluation Methods

How a Courtroom Debate Between AIs Makes Medical Claims More Reliable

Structured, courtroom-style debates with iterative, query-driven retrieval raise claim-verification accuracy by about 10 percentage points (to 81.7%), mainly because expanding evidence during debate and mixing model roles prevents premature consensus and offsets individual model errors.

Framing claim checking as an adversarial courtroom — with Plaintiff, Defense, Judge, Critic, and Exp...

Credibility:
Must Read:
Mar 30
Agent Architectures

Tell Your Car How to Drive: Personalize Its Style with Words and Your Driving History

A driving model can learn long-term user habits from recorded driving and follow short-term natural-language instructions to match individual styles, while still keeping safety and reasonable performance in simulation.

A single end-to-end model that fuses front camera views, route goals, a driver profile embedding, an...

Credibility:
Must Read:
Mar 261
Evaluation Methods

An AI helper that makes Alzheimer’s diagnosis fairer and faster

AD-CARE is an AI agent that combines whatever patient data is available (imaging, tests, genetics) to produce guideline-aligned Alzheimer’s assessments that are more consistent across centers and help clinicians diagnose faster.

AD-CARE adapts to incomplete, real-world clinical data by planning which specialized tools to run an...

Credibility:
Must Read:
Mar 26
Multi-Agent Systems

Getting Robot Teams to Track Moving Things Reliably, Even When Predictions Are Uncertain

Use learned target forecasts plus calibrated uncertainty estimates to assign and replan multi-robot tasks online, cutting average completion time and its variability while keeping task guarantees.

Combining per-target trajectory predictors with calibrated uncertainty bands lets a centralized plan...

Credibility:
Must Read:
Mar 26
Multi-Agent Systems

A Self-Improving Team That Lets Game Agents Decide Twice as Fast

A three-agent loop that prunes and compresses game observations lets large language models make real-time strategy decisions much faster and more reliably—cutting input size by about 70% and halving response time while boosting win rates.

Structural pruning of raw game observations using a graph-based entropy measure removes redundant in...

Credibility:
Must Read:
Mar 25
Multi-Agent Systems

Why Multiple Code Helpers Break When Specs Are Vague

When multiple code-writing agents work on the same class, vague specs cause a large drop in integration success; writing richer specifications largely fixes the problem.

Coordination between two code agents collapses as specification detail is removed: independent agent...

Credibility:
Must Read:
Mar 25
Evaluation Methods

Why the "Cheaper" AI Model Can Cost You More

Lower per-token prices can be misleading: invisible internal "thinking" tokens often dominate bills and can make a cheaper-listed model cost more in practice.

Marked differences in internal reasoning length — what the authors call "thinking tokens" — drive la...

Credibility:
Must Read:
Mar 25
Multi-Agent Systems

Let Big Robots Start Closer: Safer, Denser Motion for Warehouse Fleets

Reduced the required start/goal spacing for equal-sized circular robots from 4 radii to 2√3 (≈3.464), letting robots be initialized closer while keeping a provable, collision-free guarantee.

An existing continuous-space method for anonymous multi-robot routing (where any robot can go to any...

Credibility:
Must Read:
Mar 25
Multi-Agent Systems

Smart cameras that skip private spots and still watch the city well

Coordinating cameras locally lets them avoid privacy-sensitive spots by design while improving overall coverage—about 18% better coverage efficiency and an 85% reduction in privacy violations versus common baselines.

Cameras can precompute a small set of viewing options that explicitly exclude private areas and then...

Credibility:
Must Read:
Mar 24
Multi-Agent Systems

How to route many robots when they get in each other’s way

Search over which robots depend on which others (an agent dependency graph) and plan multi-step moves together — this removes the single-dependency limit of prior fast methods and improves planning when robots are large or take up more space.

Treating planning as a search over agent dependencies (who must move for whom) lets a fast priority-...

Credibility:
Must Read:
Mar 24
Tool Use & Grounding

How an AI Beat Human Teams in a Live Hacking Challenge

A protocol layer that forces AI plans into strict, executable commands greatly reduces hallucinations and enabled an autonomous agent to win a live hacking contest.

A protocol-driven design that separates high-level AI reasoning from actual tool execution prevents ...

Credibility:
Must Read:
Mar 23
Multi-Agent Systems

How cars learn to merge by watching just two neighbors

Focusing each vehicle’s decision on just the vehicle in front and the vehicle on the merging road — plus letting the model learn which past moments matter — yields faster training and much fewer collisions while keeping decentralization and low compute needs.

A simple “partial attention” setup that restricts what each car sees (only the front vehicle and the...

Credibility:
Must Read:
Mar 23
Multi-Agent Systems

How a Humanoid and a Robot Dog Find Their Way Together—No Training Needed

A humanoid coordinates and a quadruped scouts let two-robot teams navigate unseen real environments with zero training, zero prior maps, and zero simulation, reaching near human-level efficiency.

Pairing a humanoid as a coordinator with a quadruped as an explorer enables robust path planning in ...

Credibility:
Must Read:
Mar 23
Agent Architectures

A Simple Way to Make Navigation Agents Learn Faster and Safer

Blending reflex-like cues, habit learning, and goal-driven planning lets navigation agents learn faster and avoid risky areas — with a single arbitration step that picks the best strategy on the fly.

Combining Pavlovian cues (reflexive responses to contextual signals), habit-style learning, and plan...

Credibility:
Must Read:
Mar 23
Multi-Agent Systems

Fix the Question, Get Better Answers: A Simple Way to Improve AI Responses

Jointly improving how a question is asked and the model's instructions produces better AI answers; a dual-track multi-agent process lifted accuracy up to 3.95% while often using very few training examples.

Optimizing question wording and prompt instructions together yields consistently better results than...

Credibility:
Must Read:
Mar 20
Multi-Agent Systems

Satellites That Decide Fast: Onboard Teams Spot Floods and Fires Sooner

A small, role-based team of onboard agents can route work smartly so satellites raise faster, clearer alerts while avoiding wasted analysis on benign scenes.

A lightweight early-warning agent inspects incoming images and only activates specialist analysis wh...

Credibility:
Must Read:
Mar 20
Multi-Agent Systems

Robots and Home Sensors Team Up to Finish Chores Faster

Giving indoor robots a shared world model fed by home sensors (like cameras) cuts wasted searching, speeds up multi-robot tasks, and lowers the planning cost of using large language models.

A new benchmark and framework called IndoorR2X enforces realistic limited vision for each robot and ...

Credibility:
Must Read:
Mar 20
Reasoning & Planning

A Simple Checklist That Stops Web-Controlling AI From Getting Stuck

Setting and checking short, verifiable mini-goals during both runtime and training gives web-controlling AI a dense progress signal, dramatically reducing mid-task “stuck” failures and improving long-horizon success.

Automated analysis showed the dominant failure mode in web navigation is poor mid-task planning: age...

Credibility:
Must Read:
Mar 20
Multi-Agent Systems

How AI Agents Learn to Settle on Stable Strategies Without Extra Training

With two simple reasoning skills—learning other agents from history and gradually choosing better counter-strategies—language-model agents can, without additional training, settle into stable, predictable strategies along their actual play paths.

Agents that can (1) infer opponents’ strategies from observed play and (2) asymptotically learn near...

Credibility:
Must Read:
Mar 19
Multi-Agent Systems

Make a swarm cover an area in real time using far less compute

You can get exact, long-horizon coverage plans for many robots while cutting online compute growth from cubic to linear and keeping provable stability when local targets shift.

An analytic reduction collapses the long-horizon optimal control problem into a much smaller quadrat...

Credibility:
Must Read:
Mar 19
Multi-Agent Systems

How Groups Make Optimal Choices When Everyone Sees Only Part of the Picture

Representing time-stepped decision problems as a network of linked optimizations makes it possible to compute exact equilibrium strategies for linear–quadratic multi-agent problems where each agent only observes some other agents’ states.

Model each agent’s time-indexed decision problem as a node in a network that captures who sees what....

Credibility:
Must Read:
Mar 19
Multi-Agent Systems

Why Organization Rules Matter More Than Which AI You Use — Until It's Too Powerful

How authority is organized often matters more than which model you deploy for preventing abuse, but extremely capable models can overwhelm any governance design.

Simulated government-like teams of language-model agents showed that institutional setup drives corr...

Credibility:
Must Read:
Mar 19
Evaluation Methods

How AI Recommenders Can Be Tricked into Bad Choices

Even top language models used as autonomous recommenders can be reliably steered by small, realistic bias signals when choices are close — so agent recommendations need extra checks before execution.

When recommendation options are made intentionally close in quality, modern language models frequent...

Credibility:
Must Read:
Mar 18
Safety & Reliability

When Helpful Advice Makes Everyone Work Harder: Why Improving Your Score Can Become a Never‑Ending Race

Allowing everyone to follow actionable improvement steps in a ranked selection system can raise the success bar itself, turning individual gains into a collective arms race that makes upward mobility harder and can lock in inequalities.

Providing actionable directions for rejected applicants causes many people to change the same featur...

Credibility:
Must Read:
Mar 18
Multi-Agent Systems

Keep Robot Teams Safe by Fixing the Network, Not Just the Code

Proactive network coding plus a delay-aware estimator preserves localization accuracy and safety under wireless loss and delay—delay, not packet loss, breaks multi-robot coordination most.

Delayed packet delivery, not raw packet loss, is the primary driver of estimation and safety failure...

Credibility:
Must Read:
Mar 18
Multi-Agent Systems

How a Little Trust Stops Game Bots from Fighting Over Loot

A tiny, local 'trust' state attached to each agent causes self-interested agents to naturally back off, cooperate, and preserve shared resources—without changing their reward signals.

Adding a compact trust variable and short- and long-term memory to standard learners makes purely se...

Credibility:
Must Read:
Mar 18
Agent Architectures

How One Shared Memory Keeps Many AI Agents From Stepping on Each Other

A shared, governed memory layer lets dozens of autonomous agents read and write the same customer facts and policies consistently, cutting errors and enforcing compliance while preserving retrieval quality.

A dual memory approach stores both free-form insights and typed properties from the same extraction ...

Credibility:
Must Read:
Mar 18
Evaluation Methods

How AI Helpers Can Make Job Interviews Fairer and Safer

A team of specialized AI helpers coordinated by a central controller makes interview scoring more accurate, more resistant to manipulation, and easier to audit—without retraining models.

Breaking the interview workflow into four focused agents (question writer, security guard, scorer, a...

Credibility:
Must Read:
Mar 17
Agent Architectures

Stop AI Agents from Repeating Work and Cut Runtime

Pre-warm and globally reuse model prompt state across agent workflows to eliminate redundant model work—Helium speeds complex multi-step agent runs up to 1.56× while keeping exact outputs.

Treat agent workflows as data pipelines and expose prompt structure to the serving layer so the syst...

Credibility:
Must Read:
Mar 17
Multi-Agent Systems

How Drone Teams Learn to Coordinate by Talking Only to Nearby Peers

A graph-based training method lets drones learn decentralized policies that use only local sensing and short-range peer messages to reach near-optimal relay coverage and generalize to different team sizes.

A dual-attention graph encoding plus centralized training and decentralized execution produces decen...

Credibility:
Must Read:
Mar 17
Multi-Agent Systems

Stop Paying for Repeats: Cut Multi‑Agent AI Token Costs by 80%+

Only rebroadcast shared artifacts when needed (lazy invalidation) and you can cut multi-agent token transmission by roughly 80–95%, letting you run more agents and longer reasoning traces for the same budget.

Treat shared documents between agents like cached data and track simple per-agent per-artifact state...

Credibility:
Must Read:
Mar 16
Evaluation Methods

An AI That Gets Better by Challenging Itself

AI models can markedly improve math and coding accuracy by running four role-based agents—one that generates problems, one that plans, one that solves, and one that critiques—using only 500 seed examples and automatic verifiers to drive training.

A four-role setup (Challenger, Planner, Solver, Critic) lets a single model play multiple roles and ...

Credibility:
Must Read:
Mar 16
Safety & Reliability

A Tool That Finds Why Teams of AI Agents Break (and How to Watch Them)

Groups of AI agents are much less safe than individual models: automated tests across 300 workflows show an average safety pass rate of 7.1%, with system-level risks passing only 1.3%.

TrinityGuard is a unified, plug-in framework that tests and monitors safety across agent teams by tr...

Credibility:
Must Read:
Mar 16
Tool Use & Grounding

Get Safe, Local AI to Turn Your Questions into Process Insights

A two-agent local system can take plain-language business questions, generate and verify executable analysis code, and produce data-grounded reports without sending raw event logs to the cloud.

Splitting work between a code-focused agent and an interpretation-focused agent makes process analys...

Credibility:
Must Read:
Mar 16
Multi-Agent Systems

Make Robot Fleets That Steer Clear of People and Other Movers

Incorporating learned movement patterns from the environment into centralized planning cuts conflicts with uncontrolled movers, at the cost of higher planning time but without hurting overall throughput.

Using Maps of Dynamics (learned motion patterns) to bias path search makes centralized planners avoi...

Credibility:
Must Read:
Mar 13
Safety & Reliability

How to Get AI Helpers to Cooperate Without Manipulating People

High cooperation from powerful language models can be achieved through manipulation; adding a simple constitutional governance layer (hard rules plus a soft selector and dose control) preserves autonomy, fairness, and truthfulness while only modestly lowering raw cooperation.

Unconstrained language-model policy compilers push the population to the highest cooperation levels ...

Credibility:
Must Read:
Mar 13
Multi-Agent Systems

How Teaching Memory Pieces to Work Together Makes Chatbots More Personal

Jointly training the parts of a personalized memory system so they share credit for overall answers leads to better personalized responses than optimizing each part separately.

Treat the memory system as a team: a fine-grained extractor, a coarse profile builder, and a retriev...

Credibility:
Must Read:
Mar 13
Benchmarks & Datasets

Auto-draw Textbook-Quality Diagrams Quickly and Cheaply

Separate the 'what' from the 'how': having a language model list the concepts first, then using a layout engine to render them, produces large numbers of accurate, diverse diagrams cheaply.

An agent that first extracts domain knowledge as a set of conceptual ‘ideas’ and then hands those id...

Credibility:
Must Read:
Mar 13
Multi-Agent Systems

When Multiple Language Models Help — and When They Waste Time

Teams of language models speed up work only when tasks can be cleanly split up; otherwise coordination, communication, and error-handling often erase gains and add cost.

Scaling multiple language model agents follows the same limits as old-fashioned distributed computin...

Credibility:
Must Read:
Mar 12
Multi-Agent Systems

Letting Each Part Decide Alone Can Raise Your Maintenance Costs

Decentralizing control scales decision-making but can significantly worsen long‑term maintenance costs when the system has redundancy; centralized information during training helps, but coordination gaps remain. Keep centralized baselines to know when decentralization is costing you.

Decentralized multi‑agent learning works well when components form a series (every part must work), ...

Credibility:
Must Read:
Mar 12
Multi-Agent Systems

How an AI Can Write, Direct, and Film a Short Comedy—All by Itself

AI can automatically produce 1–2 minute comedy sketches that approach professional quality by running many specialized writer, director, and critic bots that compete and iteratively refine ideas based on viewer engagement data.

Multiple specialized agents—writers, critics, editors, and scene directors—can be arranged into comp...

Credibility:
Must Read:
Mar 11
Benchmarks & Datasets

One Testbed That Shows When Simple Maps Are Good Enough for Robot Fleets

Road-based abstractions (roadmaps) give nearly the same path quality as full-motion planning while being much faster; grid abstractions are fastest but raise path cost noticeably.

A single simulator lets teams run the same tasks as: raw continuous motion, a roadmap of waypoints, ...

Credibility:
Must Read:
Mar 11
Multi-Agent Systems

How Search Became a Helpful Buying Coach for Shoppers

Turning search into a proactive decision assistant—built as four cooperating AI helpers—cuts shoppers' decision effort and raises purchases, especially for complex buying situations.

A coordinated set of specialized agents (Planner, Executor, Guider, Decider) moves search from passi...

Credibility:
Must Read:
Mar 12