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

Make AI Agents Stop Overthinking — Save Time and Keep Results

Use targeted 'what-if' checks to let the runtime harness intervene only when it causally improves outcomes; this preserves or improves task success while cutting unnecessary reasoning and execution cost.

Learning the causal effect of harness interventions lets the system decide when extra reasoning or correction will actually help. A learned effect est...

Credibility:
Must Read:
Filter:

Showing 49 papers

Safety & Reliability

How AI Agents Can Prove Who They Are and Who's Responsible

InterSAGE makes agent-to-agent collaboration verifiable by cryptographically binding identity, advertised skills, delegation limits, and signed execution logs so agents cannot impersonate, escalate authority, or repudiate actions.

A four-layer trust substrate (identity, discovery, trust negotiation, accountability) lets agents ad...

Credibility:
Must Read:
Aug 13
Multi-Agent Systems

Teach Multiple Agents What to Share so They Act Better Together

Jointly designing what agents communicate and how they act can be solved in many linear, noisy systems: under mild structural rules, optimal controllers remain simple (linear) and can be computed with closed-form equations.

When agents operate in a linear system with quadratic costs and Gaussian noise, properly restricting...

Credibility:
Must Read:
Aug 13
Multi-Agent Systems

Make diverse smart devices stay in sync without sending every detail

Translate and send small, targeted belief updates via nearby servers so heterogeneous AI agents share a common understanding with far less bandwidth and much lower error than naive matching-only syncing.

Sending compact belief updates (not full models) to local servers that translate them into each rece...

Credibility:
Must Read:
Aug 13
Multi-Agent Systems

Getting Hundreds of Robots to Find a Hidden Exit Without Colliding

A simple leader‑and‑followers scheme using only short‑range messages lets hundreds of agents navigate unknown maze‑like graphs to a hidden goal with provable safety (no collisions) and time performance that approaches an optimal strategy that knows the map in advance.

A single active agent (the head) runs a normal single‑agent maze solver while all other agents form ...

Credibility:
Must Read:
Aug 12
Multi-Agent Systems

Telling AIs They're Alike Makes Them Cooperate — Sometimes

Higher reported similarity between AI agents usually makes them act more cooperatively, but the effect varies widely by model and depends critically on how similarity is measured and shared.

Providing a numeric similarity score between two large language model agents often increases coopera...

Credibility:
Must Read:
Aug 12
Evaluation Methods

How AI Can Act Like a Lab to Explain and Fix Other AI

An automated, multi-agent system can use AI itself as a scientific instrument to discover how other AI models work, find hidden risks, and produce targeted fixes that measurably improve behavior. The process combines a large evidence base, a library of mechanistic methods, and iterative experiment checks to generate reliable, testable insights.

An agentic system named Mechanist generates hypotheses about model mechanisms, runs experiments, ver...

Credibility:
Must Read:
Aug 12
Safety & Reliability

Make AI Assistants Safer by Controlling What They Send Out

Enforce privacy and security for AI agents at the network level by pairing per-agent mediators with a central policy plane: combine deterministic rules with context-aware checks so agents can’t unknowingly leak or misuse data.

Security works better when enforcement sits outside the agent and inspects every external interactio...

Credibility:
Must Read:
Aug 12
Evaluation Methods

How Small AI Agents Help Find and Fix Sensor Problems

Splitting reasoning into small, specialized AI agents that work with a digital twin makes anomaly diagnosis and mitigation more interpretable and lets lightweight open models match or beat larger monolithic setups.

A digital twin generates expected sensor values and an anomaly classifier creates structured events ...

Credibility:
Must Read:
Aug 12
Evaluation Methods

Turn a Plain Research Request into a Shareable, Run-Ready Analysis Plan

AI can turn natural-language research descriptions into visual, executable analysis workflows that run reliably and can be edited cheaply instead of rebuilt.

Natural-language task descriptions were automatically converted into deterministic, inspectable work...

Credibility:
Must Read:
Aug 11
Multi-Agent Systems

How future mobile networks can self-organize using conversational domain helpers

Local conversational agents at each administrative domain can coordinate end-to-end service setup across a changing multi-domain mobile network by exchanging compact, neighbor-only reachability summaries—keeping coordination cheap, simple, and plug-and-play.

Deploy a conversational agent inside each administrative domain to summarize local resources, share ...

Credibility:
Must Read:
Aug 11
Evaluation Methods

Find Which AI Helper Broke the Plan — Faster, Clearer Debugging for Teams of AI

ASCon uses step-by-step and agent-level context together to accurately identify the responsible agent, the faulty steps, and the failure mode — boosting attribution accuracy by double-digit percentages in key metrics.

Modeling the execution trace as two directed graphs (one over steps, one over agents) and letting st...

Credibility:
Must Read:
Aug 11
Human Oversight

Make AI Explanations That Match the Person Reading Them

A small multi-agent pipeline that fixes the model evidence and then plans, writes, and reviews explanations produces audience-specific explanations that are clearer, less templated, and remain faithful to the underlying attributions.

A structured ExplanationCard holds a single, immutable summary of the prediction and feature attribu...

Credibility:
Must Read:
Aug 11
Safety & Reliability

How Clear Rules and Approval Records Stop AI Helpers from Sharing Sensitive Files

Identical safety numbers can hide very different realities: written rules can stop unsafe proposals before they happen, while code-only guards may block execution but still let risky proposals appear; provenance-aware guards both prevent unsafe proposals and let work continue safely.

When agents decide where to send work, lowering the attractiveness of a safe internal route makes un...

Credibility:
Must Read:
Aug 10
Multi-Agent Systems

Teach a Crowd to Share: Learn Rewards That Stop Resource Collapse

Learning a shared reward model from episode-level judgments lets many independent agents develop cooperative norms and avoid resource collapse, even when standard policy training repeatedly fails.

A shared reward predictor trained on whole-episode social evaluations can turn sparse, global object...

Credibility:
Must Read:
Aug 7
Reasoning & Planning

An AI that designs lab recipes for hard-to-make natural molecules

A multi-agent AI system named SynthEx can design complete synthetic routes for 63.9% of challenging natural products (vs 13.8% for a leading template-based planner) by generating atom-level reaction edits and planning strategy-first.

Planning strategy before individual steps and letting a language model write atom-by-atom reaction e...

Credibility:
Must Read:
Aug 7
Evaluation Methods

Why Per-Ship Personalization Didn’t Help Predict Arctic Ship Routes

Once a reward model can represent nonlinear effects of route, vessel, and environment, adding per-ship latent personalization does not improve prediction, route generation, or reward transfer.

A controlled study of 3,186 Arctic voyages across nine seasons compared three reward-learning approa...

Credibility:
Must Read:
Aug 6
Safety & Reliability

Let People Vote on How Much Compute an AI Can Use — and Lock It in Hardware

Make compute the governance lever: stakeholders cast breadth-weighted votes that convert into a signed, hardware-enforced compute license, giving communities a continuous, accountable stop/go lever while capping deployer liability.

A two-sided, participatory market combines supporters and objectors using a breadth-weighted aggrega...

Credibility:
Must Read:
Aug 6
Agent Architectures

How Hospitals Can Scale AI Safely and Cut Administrative Work

A compliance-first, multi-layer AI platform that coordinates multiple specialized agents and embeds policy-as-code can turn siloed hospital pilots into scalable, auditable systems that cut duplication and reduce manual work.

A layered architecture—agent orchestration, centralized compliance and policy-as-code, plus a privac...

Credibility:
Must Read:
Aug 61
Agent Architectures

How Your Network Can Help AI Agents Finish More Tasks Without More Hardware

Steering agent calls with temporary, revisable network reservations lets more agent workflows meet deadlines using existing capacity—tested gains up to 10% more completed workflow value.

Many AI-agent workflows reveal at least one remote call before the connection is opened, giving the ...

Credibility:
Must Read:
Aug 6
Evaluation Methods

Stop Testing Agents Too Long — Save 74× on Evaluation Costs

Combine variance-aware payoff corrections with anytime-valid stopping rules to safely stop interactive agent tests much earlier — about 74× fewer hands in realistic poker agent runs — while keeping claims auditable.

Merging a variance-reducing correction for each game with a time-uniform confidence sequence makes c...

Credibility:
Must Read:
Aug 6
Multi-Agent Systems

Let Many AI Workers Try Different Ideas — They Find Better Solutions

A shared withheld-test leaderboard plus many independent AI agents searching in parallel produces broader, qualitatively different solutions than a single agent repeatedly refining one approach, while still meeting high precision requirements.

Using a withheld-test leaderboard as the only shared reward, multiple autonomous agents running in p...

Credibility:
Must Read:
Aug 5
Multi-Agent Systems

How Multiple AI Agents Solve Hard Problems Without Talking Too Much

Build a union of diverse agent team layouts, but only let small, trusted pieces talk at a time and locally fix risky parts — that cuts distracting noise and improves final answers (avg +3.47%, up to +10.34%).

Selecting a small set of complementary collaboration patterns and merging them into a union keeps di...

Credibility:
Must Read:
Aug 5
Multi-Agent Systems

Why AI Helpers Still Fall for Freeloaders (and How to Spot It)

LLM-based agents naturally learn simple, gossip-like reputation rules that reward past generosity, but they rarely develop the deeper norms needed to reliably resist free-riders—best models expelled defectors only about half the time.

Agents that watch the last three interactions and choose donations from 0–100% evolve threshold-styl...

Credibility:
Must Read:
Aug 5
Human Oversight

A Care Plan You Can Question: Let People Fix AI's Home Care Suggestions

Make AI care recommendations inspectable and contestable so clinicians and caregivers can correct, reject, or improve plans before they reach the patient.

CoPlan turns AI recommendations into editable proposals by having multiple specialized AI agents gen...

Credibility:
Must Read:
Aug 5
Multi-Agent Systems

Let Cheaper Agents Handle the Easy Stuff: Save Costs Without Losing Accuracy

History-aware delegation and a simple virtual money system let a group of differently skilled agents hand off routine work to cheaper peers while reserving expensive agents for hard steps, cutting operating cost without dropping task performance.

Treating delegation as a learned decision lets a coordinator assign each step to the most cost-effec...

Credibility:
Must Read:
Aug 4
Evaluation Methods

Why AI Agents Lose Your Instructions in Other Languages — and an Easy Fix

Multilingual requests often lose critical constraints when turned into plans; representing five explicit commitments before planning recovers meaningful accuracy across languages and models.

Five repeatable ways plans lose track of user intent were identified: wrong entity, wrong source, wr...

Credibility:
Must Read:
Aug 4
Safety & Reliability

How AI Can Weigh Different Human Views Without Picking a Single ‘Right’ Answer

Pluralistic alignment works when AI models treat perspectives as social roles that deliberate and are aggregated with attention to power and provenance, and when alignment is judged over interaction histories rather than one-off answers.

Social theory gives concrete tools to make agentic AI represent and coordinate multiple legitimate p...

Credibility:
Must Read:
Aug 4
Multi-Agent Systems

How letting AI agents pick who to listen to stops groupthink and uses fewer words

Regulating who each agent listens to during a group debate reduces blind conformity, improves final answers across reasoning and question-answering tasks, and lowers total token use.

Looking at the debate as a group process — not just evaluating each agent alone — uncovers when the ...

Credibility:
Must Read:
Aug 4
Safety & Reliability

How to Keep Autonomous AIs From Going Off Script

Security must move beyond checking single actions to verifying an agent's behavior over time and across interacting components—covering tool integrity, inter-agent trust, model routing, and supply chains.

Agentic systems face four broad, linked risks: untrusted inputs that hijack reasoning, persistent me...

Credibility:
Must Read:
Aug 3
Evaluation Methods

When Simple Code Search Beats Complex AI Teams

Index-backed retrieval (simple semantic search) answered more repository questions correctly and at lower cost than delegation-based multi-agent exploration for read-only, indexable codebases.

Across 15 Python repositories and four large language models, a retrieval agent that uses a pre-buil...

Credibility:
Must Read:
Aug 2
Evaluation Methods

One strategy that keeps teams both safe and productive — but it can be costly

A single joint strategy can be required to both meet temporal goals (like “serve both bays forever”) and guarantee long‑run average rewards, but verifying and implementing such strategies can demand large computation and nontrivial strategy memory.

A logic was defined that combines temporal requirements with long‑run average (mean) rewards so a co...

Credibility:
Must Read:
Aug 2
Multi-Agent Systems

How to Test What Makes Groups of AI Agents Trust—or Betray—Each Other

A replay-validated rules engine makes it possible to hold game mechanics constant while varying communication and command styles, so you can reliably compare how those factors change multi-agent decisions and audit every outcome. The WOPR environment proved engine parity across 240 seeded games, so results are reproducible and replayable.

Built a deterministic simulation environment (WOPR) that enforces a fixed rulebook, records every st...

Credibility:
Must Read:
Aug 3
Multi-Agent Systems

A single reply-randomness dial that controls whether chat agents agree

Two simple listener behaviors — how often an agent accepts a known word and how often it accepts a new word — predict whether a group of chat agents will converge on a single name; changing reply randomness moves those rates and can slow or speed consensus by several times.

Measuring two conditional rates (accept-known-word π and accept-new-word ϕ) compresses the messy beh...

Credibility:
Must Read:
Aug 3
Evaluation Methods

How to Trust an AI That Makes Decisions Over Time

Validate whole decision paths, not just single outputs: outcome-only tests miss trajectory-level failures that produce real-world harm, and current literature leaves lifecycle monitoring and evidence trails largely unaddressed.

Most evaluation efforts measure whether an agent reaches a correct end result, but many failures occ...

Credibility:
Must Read:
Jul 31
Multi-Agent Systems

How teams of AI helpers can fix their own mistakes without a person digging through logs

Automating repairs by exploring rollback-and-repair choices leads to reliably fixing more multi-agent failures (6–12% higher success) while keeping API/token costs low.

Searching a space of rollback-and-repair moves finds better fixes than linear or single-path methods...

Credibility:
Must Read:
Jul 31
Agent Architectures

Turn a Few Past Snapshots into Long-Term Memory for Smarter Conversations

TransMem converts a small set of past internal model snapshots into a compact, task-aware memory so frozen language models can reason over very long conversations without bigger models or retraining.

Hidden states (internal snapshots of a model’s computation) already contain compact, position-sensit...

Credibility:
Must Read:
Jul 31
Multi-Agent Systems

An AI Team That Helps Neuroscientists Find Brain Insights Faster

A specialized AI multi-agent system can plan and run complex, cross-modal neuroscience analyses, cutting routine work and producing valid discoveries—it beat general coding agents by up to 17.7% on a domain benchmark.

SeekBrain builds a growing library of field-tested analysis recipes pulled from papers and open code...

Credibility:
Must Read:
Jul 31
Multi-Agent Systems

When to Watch Others and When to Go It Alone

People choose to observe others when simulating their goals and beliefs promises more useful information than acting now; a model that combines mind-reading with cost–benefit reasoning predicts human observation choices closely.

Humans balance the benefit of watching another agent (what that agent’s actions reveal about the wor...

Credibility:
Must Read:
Jul 30
Multi-Agent Systems

Letting Coding AIs Hear Each Other Without Stopping Doubles Success

Background listening between working coding agents—so messages surface between steps without interrupting work—nearly doubles task success on long software-understanding problems, while requiring only a small harness change.

Adding a lightweight message layer where agents can listen in the background (passive awareness) let...

Credibility:
Must Read:
Jul 30
Safety & Reliability

How an Agent Can Safely Profit from a Weak Opponent

Agents can safely exploit non-optimal opponents by certifying each exploit’s downside before playing: you get large gains when the opponent is exploitable, and you never exceed a preset loss budget.

A practical online method (CS-RNR) watches opponent action frequencies with time-safe confidence che...

Credibility:
Must Read:
Jul 30
Evaluation Methods

Don't Judge an Idea by One Run: Re-running Code Can Flip Research Picks

Single-run results often misrepresent an idea: 25–44% of one-run winners reverse when the same idea is independently reimplemented, so automated research should test multiple realizations before crediting an idea.

Implementation choices (the ordinary coding decisions left open by a research idea) create far more ...

Credibility:
Must Read:
Jul 29
Multi-Agent Systems

Teach AI What a Teammate Can Actually Do — So It Stops Asking for the Impossible

Estimating a partner’s stable capabilities from a few joint tasks lets an AI avoid assigning infeasible work and reuse that knowledge across new tasks — but you must handle diverse, suboptimal human behaviour for robust results.

Learning an explicit, task‑independent model of what a partner can and cannot do (their capabilities...

Credibility:
Must Read:
Jul 29
Safety & Reliability

Teach Your AI System to Fix Its Own Repeated Mistakes

Convert evaluated failures into persistent procedure fixes by evolving an external harness state (episodic memory + state graph), and you can boost task success by about 10 percentage points without changing the base model or tools.

Living-Harness turns after-action evaluations into bounded, reusable updates to an external harness ...

Credibility:
Must Read:
Jul 29
Evaluation Methods

Teach your AI agent to improve itself — 16% better with fewer wasted tests

Reusing past runs smarter (by tying evidence to specific functions and recalibrating whether that evidence still applies) makes automated harness updates steadier and yields about 16% better reasoning and 14% better agent-task performance under limited evaluation budgets.

Historical trial logs often mislead automated harness updates because evidence becomes stale or lack...

Credibility:
Must Read:
Jul 29
Multi-Agent Systems

Turn Building Rules into Automatic, Inspectable Code

A structured multi-agent, test-driven harness converts expert-written rule interpretations and labeled models into executable, inspectable compliance checkers, boosting average element-level accuracy from 46.8% to 84.98%.

A deterministic multi-agent loop (planner, generator, evaluator) that follows test-driven developmen...

Credibility:
Must Read:
Jul 28
Evaluation Methods

How Large Language Models Help Find 13 Hidden Bugs in Quantum Libraries

Combining a code-aware knowledge base with large language model test generation and quantum-aware mutations produces many more valid tests and finds real faults — KQFuzz uncovered 13 bugs (12 fixed) and boosted coverage by up to 18.44%.

A knowledge-guided approach that feeds library-specific API facts into a language model produces far...

Credibility:
Must Read:
Jul 28
Agent Architectures

Stop AI Agents From Stepping on Each Other in the Cloud

Make coordination part of the cloud system, not something AI agents must guess. CloudWeaver ties each agent session to an authoritative shared view so independent changes run in parallel while semantic conflicts are prevented.

AI agents and human teams operate on shared cloud resources from limited, session-scoped views; when...

Credibility:
Must Read:
Jul 28
Tool Use & Grounding

How a virtual analyst makes brain scans easier to understand

Combining a language model with topology-aware graph tools, targeted neuroscience retrieval, and a reflection step yields more accurate, less overconfident, and better-explained brain-graph predictions.

BrainAgent turns a brain connectivity graph into compact, multi-level descriptions (region, subgraph...

Credibility:
Must Read:
Jul 24
Multi-Agent Systems

Why Independent AI Agents Can Drain a Shared Energy Pool (and What That Means)

Independent AI agents that only protect their own service can collectively deplete a shared renewable reserve when demand outpaces renewal, producing costly fallback and long-term capacity loss even when sustaining use was feasible.

Four AI agents acting as electricity prosumers, each trying only to keep their own power running and...

Credibility:
Must Read:
Jul 24