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

How AI Can Cut Wasted Computation by Over Half

Plugging language-model reasoning into a fixed orchestration architecture can cut wasted computation by up to 55% and reduce GPU busy time by about 40% for certain scientific campaigns.

A simple, three-part architecture (one component for logical decisions, one for running tasks, and one for tracking evidence) makes it easy to swap in...

Credibility:
Must Read:
Filter:

Showing 49 papers

Multi-Agent Systems

How robot teams avoid hidden obstacles without getting stuck

Sharing a spatial belief about which spots might be blocked and turning that belief into detour-aware route costs lets planners pick safer, cheaper routes—cutting real-world travel cost and failures even at large scale.

Agents that maintain a shared probabilistic map of uncertain locations and update it from local obse...

Credibility:
Must Read:
Sep 30
Multi-Agent Systems

How Tens of Robots Can Move Boxes Together Without Getting Stuck

A configuration-based planner adapted from multi-robot pathfinding lets teams of agents coordinate pushes and solve multi-box moving tasks quickly, reliably, and often optimally—scaling beyond single-agent search methods.

Adapting a configuration-space search backbone plus a coordinated action generator produces a solver...

Credibility:
Must Read:
Sep 30
Evaluation Methods

Check Commands Before Running Them — Make AI Agents 36% More Likely to Finish Tasks

Verifying multiple candidate actions before executing any command can turn the same model into a far more reliable agent — a strong verifier raised task success from 50% to 68% without changing the generator.

Sampling several candidate commands from a fixed model creates useful alternatives, but most of that...

Credibility:
Must Read:
Sep 30
Multi-Agent Systems

Train Site-Specific Sensing and Wireless AI Without Rebuilding the World

An agent-driven workflow turns a plain-language deployment request into a coordinated sim-to-real plan that builds physically aligned synthetic sensing and wireless data, enabling zero-shot and few-shot transfer for vehicle detection and beam prediction while cutting repeated real-world data collection.

A two-agent setup (scene construction and scene understanding) can interpret deployment needs, plan ...

Credibility:
Must Read:
Sep 30
Multi-Agent Systems

Teach a team of AI agents new rules on the fly — no retraining needed

A simple per-agent guidance layer makes pretrained multi-agent generative models satisfy new hard constraints at generation time, eliminating retraining and fixing failures while keeping decentralization.

Pretrained multi-agent generative models can be steered at generation time by letting each agent com...

Credibility:
Must Read:
Sep 29
Benchmarks & Datasets

Train many bidding AIs fast and see why they miss big profits in power markets

A GPU-ready benchmark makes training many market-bidding agents tens of times faster and shows independent learners often fail to discover profitable coordinated strategies—so market rules shape what agents can learn.

A unified JAX benchmark implements five realistic power markets and runs market clearing and agent l...

Credibility:
Must Read:
Sep 29
Multi-Agent Systems

How Two-Bit Messages Let Teamed Agents Make Better Decisions

Two bits per message, when interpreted by each receiver using its own local context, preserve the decision-relevant differences receivers need and often match or outperform methods that send thousands of bits.

Messages should preserve the receiver’s differences in action value within the receiver’s own situat...

Credibility:
Must Read:
Sep 29
Human Oversight

When Your Assistant Should Ask a Question — and When It Should Just Do It

Estimating the expected benefit of a question (its value-of-information) lets an assistant ask far fewer clarifications while producing more correct, preference-aligned outcomes.

A decision rule that weighs the expected improvement in task reward against the cost of asking leads...

Credibility:
Must Read:
Sep 29
Multi-Agent Systems

How robot teams plan safe paths fast

Iterative noise-removal sampling (a diffusion denoising approach) produces fast, scalable, and reliable multi-robot paths that work in simulation and real hardware.

Diffusion denoising—an approach that repeatedly refines many noisy candidate control sequences—lets ...

Credibility:
Must Read:
Sep 28
Multi-Agent Systems

When AI Helpers Swap Strategies They Solve Harder Problems

Let the best-performing agent share its strategy and rebuild who listens to whom for each problem — teams of AI agents reason better without any extra training or external judge.

Selecting a single agent as a temporary “strategy donor” based on early answers, copying its role in...

Credibility:
Must Read:
Sep 28
Multi-Agent Systems

Keeping the Same Trading Partner Makes Pricing Bots Charge More

Agents that keep the same partner learn to set higher prices; randomly rematching counterparties reduces those learned prices even when no punishment or messages are possible.

When two pricing agents repeatedly face the same rival, the prices they learn are higher than when t...

Credibility:
Must Read:
Sep 28
Safety & Reliability

Stop Fake Agreement: Make AI Teams Admit When They Disagree

Enforce a post-debate verification gate that requires every claim to be tied to the debate record; doing so raises groundedness and either produces verifiable conclusions or clearly signals disagreement.

When you force the final synthesizer to justify each claim against a frozen debate record, many flue...

Credibility:
Must Read:
Sep 25
Benchmarks & Datasets

When AI Agent Teams Only Finish Half Their Tasks

Top large language models working as teams complete only about half of long, multi-step tasks, and fewer than one in three agent actions actually help reach the goal.

AgentWorld is a long-horizon simulator built to test how groups of AI agents coordinate over 25–55 r...

Credibility:
Must Read:
Sep 25
Evaluation Methods

Why Adding More AI Helpers Often Stops Improving Results — and When Smaller, Mixed Teams Win

On tasks where one correct answer matters, larger teams hit a voting ceiling tied to the model’s most common answer; on numeric estimation tasks, simple averaging barely helps because item-level bias dominates, so diversity or better base models matter more than scale.

Scaling behaves very differently depending on the task type. For multiple-choice or problem-solving ...

Credibility:
Must Read:
Sep 25
Safety & Reliability

How AI Agents Found Each Other, Built a Community, and Skipped Oversight

Autonomous agents will improvise social networks on whatever channel survives, creating identities and norms that enable coordination but can still hold wrong beliefs or act beyond their authorization—so agent-to-agent trust needs different checks than simple interaction guarantees.

Two separate 2026 incidents show groups of deployed agents with no approved way to coordinate noneth...

Credibility:
Must Read:
Sep 25
Evaluation Methods

Let the Rules Sign Off: Make AI Agents Prove They Did the Job

Treat agents as proposers, not judges: let an external runtime convert specs into obligations and commit state only from authorized evidence so agents can’t falsely claim completion.

Agents often understand requirements but fail to execute them correctly, and they frequently declare...

Credibility:
Must Read:
Sep 24
Multi-Agent Systems

Match People to Things Fast by Focusing on What Matters Most

Projecting multi-attribute items onto their single strongest common direction often produces fast, deterministic matchings that are near-optimal in welfare — but the guarantee only holds when one feature component clearly dominates and agents can’t game the reports.

Reducing multi-dimensional preferences to a single score using a spectral (singular value) projectio...

Credibility:
Must Read:
Sep 24
Evaluation Methods

How AI Can Make Reliable Step-by-Step Plans

Split planning into three isolated steps—generate a blueprint, locally revise strategies, then independently verify—and standard language models produce far more reliable multi-step plans (up to 16.7% absolute accuracy gain).

Separating plan generation, localized revision, and independent verification prevents the error casc...

Credibility:
Must Read:
Sep 24
Evaluation Methods

How to Rehearse Public Reaction Before You Ship a Change

How you ask an evaluation matters more than which model you use: prompt and evaluation design can create large apparent gaps between models, and a simulated reaction system can recover most real public concerns but can also amplify pessimistic bias.

A synthetic lab that turns a proposed change into a knowledge graph, populates simulated personas, a...

Credibility:
Must Read:
Sep 24
Evaluation Methods

A Simple Score Can Pick the Wrong Robot Team — Losing Up to 43% of the Map

Keeping only single-robot and pairwise interaction terms can make your selector choose plans that deliver far less coverage — sometimes losing over 40% of the map. Measure the delivered outcome of the selected plan, not just how well your score fits the data.

Approximating the delivered-coverage objective by keeping only singleton and pairwise interaction te...

Credibility:
Must Read:
Sep 24
Multi-Agent Systems

How simple robots can spread out without breaking their links

Each agent can use just its farthest visible distance to move up to half the remaining visibility gap in any direction and guarantee no current link is lost, so a connected swarm stays connected while it spreads.

A single scalar — the farthest visible neighbor distance — certifies the largest safe isotropic move...

Credibility:
Must Read:
Sep 23
Safety & Reliability

How banks can stop hidden bias when multiple AI systems work together

Local checks on individual decision models can still produce biased institutional outcomes when those models interact; fixing that requires population-level monitoring, bounded authority, and runtime containment across the whole workflow.

Local fairness or safety checks on each model do not guarantee fair or compliant outcomes once multi...

Credibility:
Must Read:
Sep 23
Multi-Agent Systems

A Tiny Adapter That Lets Solo Agent Skills Work in Teams

You can reuse an already-trained single-agent controller unchanged and add a small input adapter to handle teammates, updating only 3.5–7.3% of parameters while matching or beating multi-agent training from scratch.

A compact input-side adapter translates multi-agent observations into the format expected by a froze...

Credibility:
Must Read:
Sep 22
Evaluation Methods

Asking Agents What They Expect Makes Them Cooperate More

Privately asking each agent about what others do and what should be done raises cooperation and reveals whether group behavior rests on shared beliefs or simple imitation.

Prompting agents to report their expectations consistently increased contributions across four model...

Credibility:
Must Read:
Sep 22
Evaluation Methods

When AI Helpers Agree, You Keep Nearly Twice as Many True Facts

Using agreement across diverse agents plus per-example calibration reliably controls factual errors and almost doubles retained-claim precision; extra counterfactual checks only help when the verifier actually knows the domain.

Decomposing long answers into atomic claims and ranking those claims by how many different agents se...

Credibility:
Must Read:
Sep 22
Safety & Reliability

How smart helpers can keep dementia care on track

A governed agent system can reliably turn scattered observations into tracked, permissioned care tasks and verifiable outcomes—passing an 18-trace architecture test with no policy violations.

A runtime that separates planning from enforceable authority, keeps versioned memory, and requires h...

Credibility:
Must Read:
Sep 22
Multi-Agent Systems

How 1,024 Digital Helpers Team Up to Rebuild Huge Software Faster

More cooperating workers steadily improves both speed and final accuracy: increasing worker count is a practical lever to get better results on long, complex software tasks without a central manager.

A shared workspace and simple, self-driven loop let many independent workers coordinate without a ce...

Credibility:
Must Read:
Sep 22
Multi-Agent Systems

A Simple Trick That Keeps Robot Teams Connected While They Move

A flow-based connectivity formulation lets a centralized planner keep robot teams connected while they move, using far fewer constraints and enabling better mission outcomes under the same compute budget.

Flow-based connectivity encodes exactly the conditions needed for a network to be connected, and it ...

Credibility:
Must Read:
Sep 21
Multi-Agent Systems

How Your Watch Data Can Give Clear Health Answers Without the Confusion

Break complex wearable queries into separate user requests and small, typed tasks; a dedicated query agent then pulls only the exact evidence each task needs, keeping answers traceable while using much less context and improving trust and transparency.

Decompose a multi-part wearable question into distinct intents and ordered tasks so each sub-request...

Credibility:
Must Read:
Sep 21
Evaluation Methods

How to Tell If a Market Simulator Really Matches Real Customers

Recovering model parameters is not enough — run reachability tests and held-out-statistic audits because a simulator can fit parameters yet still fail to reproduce key market outcomes like average purchase level.

An adequacy-aware calibration routine showed behavioral parameters could be recovered across four ma...

Credibility:
Must Read:
Sep 21
Multi-Agent Systems

How drone teams share clear images fast so AI can spot targets in ~21 ms

Treating whole images as the communication unit — with concurrent sending, priority-aware queues, and a cheap image-quality filter — cuts end-to-end object-detection time to ~21 ms on average and preserves detection quality under noisy conditions.

A middleware that fragments and reconstructs full images, sends fragments concurrently, prioritizes ...

Credibility:
Must Read:
Sep 21
Multi-Agent Systems

One Simple Dial That Controls How AI Groups Change Their Minds

Adding a small Bayesian belief layer outside each language model gives a single per-agent knob that reliably controls how much agents change their minds and makes those changes auditable and recoverable after the AI speaks.

A compact belief layer stores each agent’s stance as a probability and updates it by a single Bayesi...

Credibility:
Must Read:
Sep 18
Evaluation Methods

Test neighborhood energy plans before they go live

CityLearn v3 lets teams simulate realistic neighborhood energy communities—including membership changes, equipment failures and local cost settlements—while recording both what controllers requested and what was actually applied so decisions can be fairly compared and audited.

A configurable simulator and evaluation framework preserves scenario rules (members, device limits, ...

Credibility:
Must Read:
Sep 18
Safety & Reliability

How to Turn an AI Assistant Idea into a Safe, Working Service

AI-GRACE is a practical checklist that converts business goals and legal obligations into concrete technical requirements and an authorized operating envelope so teams can decide if and how to safely deploy AI assistants.

AI-GRACE links objectives, obligations, and risk to the specific capabilities and architecture neede...

Credibility:
Must Read:
Sep 18
Evaluation Methods

AI groups agree 34–44 points more than people — often on the wrong answer

Simulated AI agent groups produce much higher rates of full agreement than matched human groups (by roughly 34–44 percentage points), but that agreement often ignores correctness, so consensus is a weak trust signal.

When replaying 100 human discussion groups on a logic task with matched AI agent groups seeded with ...

Credibility:
Must Read:
Sep 17
Safety & Reliability

Make Connected-Car Safety Tests Realistic Without Rewriting Radio Code

A compact, explainable five-stage model reproduces 5G sidelink message losses measured in a detailed reference simulator, letting teams keep their existing traffic/network stack while getting realistic delivery behavior.

A distilled model called NS3Learn captures the dominant causes of message loss in 5G sidelink Mode-2...

Credibility:
Must Read:
Sep 17
Safety & Reliability

A Little Randomness Lets Two Agents Avoid Bad Outcomes

Allowing each agent to randomly pick among near-best actions (with limited randomness) guarantees the system will keep visiting outcomes strictly better than half of optimal, and makes arbitrarily bad states only temporary.

Introducing a controlled “noise” that lets agents occasionally choose near-best (not strictly best) ...

Credibility:
Must Read:
Sep 10
Multi-Agent Systems

Organize Robot Teams Like Real Crews and They Complete Wildfire Jobs Faster

Match the team structure to the task: grouping agents for parallel work and stacking managers for ordered steps makes heterogeneous robot teams complete complex missions faster, with better outcomes and competitive compute use.

Task-specific organizational hierarchies that mirror how work is interdependent (parallel vs. sequen...

Credibility:
Must Read:
Sep 10
Evaluation Methods

Turn Years of Messy Lab Reports into a Single Searchable Memory

A two-layer graph (one that preserves every document chunk and one that encodes domain concepts) plus a router that picks the right agent gives much better, traceable answers than flat document search alone.

A lossless lexical layer keeps every document, section, table, and figure retrievable for exact look...

Credibility:
Must Read:
Sep 10
Multi-Agent Systems

Why Money Gets Stuck in a Simulated Town: AI Agents Earn but Rarely Spend

Demand lifts business revenue sharply, but the cash stays with firms or in agent accounts rather than flowing to workers or customers; direct transfers are mostly saved, so short runs make the economy look frozen even though it slowly changes over longer horizons.

In a fully accounted 100-agent simulation of a real town with 762 businesses, a tourist surge raised...

Credibility:
Must Read:
Sep 10
Evaluation Methods

A Simple Second Opinion That Stops AI Agents from Making the Same Mistake

A label-free "reverse" inference — a Bayesian-style second opinion computed from the evidence in a different direction — provides a reliable anchor that improves how multiple AI agents are combined, reducing group errors especially when agents disagree.

Forward-only aggregation (voting or a single judge) can echo shared mistakes because every agent rea...

Credibility:
Must Read:
Sep 10
Multi-Agent Systems

How big warehouse robots get through narrow aisles without stopping production

A centralized traffic manager that periodically replans paths, allocates real-time movement permissions, and actively detects and resolves deadlocks lets large, different robots navigate narrow industrial layouts reliably and with higher throughput than traditional rule-based systems.

A two-layer environment model (a detailed roadmap plus topological sectors) plus a lifelong multi-ag...

Credibility:
Must Read:
Sep 9
Agent Architectures

How a Two-Part System Tags Sensitive Columns Without Seeing Your Data

A two-agent system can generate human-readable column descriptions and reliably tag sensitive columns across millions of enterprise fields without ever reading cell values, by fusing metadata, rules, and model reasoning to prioritize recall and auditability.

Combining three value-free signals — a metadata example retriever, rule/regex matches, and an LLM-ba...

Credibility:
Must Read:
Sep 9
Human Oversight

Turn Vague Notes Into Fixable To‑Dos for Animation Teams

Keep a supervisor's intent, the reference evidence, and a prioritized action list together so feedback is immediately actionable—leading to fewer clarifying questions and clearer tasks for junior artists.

Linking a supervisor’s goal to the exact reference region and to an executor-ready checklist makes r...

Credibility:
Must Read:
Sep 9
Multi-Agent Systems

One wiki note that could steer thousands of AI bots

AI runs on a public wiki mostly copied what they could see, and that simple copying explains how they coordinated — which also makes them easy to steer by whoever writes first.

Agents arriving with no memory and a browser made three linked choices — where to write, what name t...

Credibility:
Must Read:
Sep 8
Multi-Agent Systems

How Risky or Cautious Agents Push a Group Toward a Greener or Safer Choice

Agents' risk attitude during learning steers which coordination outcome the group settles on: being more risk-seeking makes the group converge to the higher-payoff (but riskier) equilibrium, while being more risk-averse steers them to the safer maximin equilibrium; a single risk threshold marks the switch.

Evaluating payoffs with an entropic risk measure (which tilts value toward favorable or unfavorable ...

Credibility:
Must Read:
Sep 8
Multi-Agent Systems

Analyze Organoid Images Automatically and Get Ready-to-Use Reports

A coordinated team of foundation-model assistants can segment organoids without any manual labeling, compute key morphology measurements, and produce a complete, reproducible analysis report from a plain-language request.

MorphoOrgaAgent turns a natural-language question from a biologist into a full analysis pipeline: it...

Credibility:
Must Read:
Sep 8
Agent Architectures

A Simple, Editable Map That Helps AI Agents Know What to Do Next

Representing procedural knowledge as an explicit, editable graph gives language-model agents clear, situation-aware guidance and reliably improves performance across diverse tasks; the graph can also be refined automatically from experience.

Organize steps, actions, and states as a directed graph where each edge carries when-to-act notes, h...

Credibility:
Must Read:
Sep 8
Multi-Agent Systems

Treat Changing Data Like a Story: A Handy Framework for Time-Based Signals

Model time-varying data as a single abstract “story” and you get a unified way to compare representations, measure what’s lost when you change formats, and reason about systems whose connections change over time.

A unified framework called narratives treats any time-varying object (signals, graphs, control traje...

Credibility:
Must Read:
Sep 8