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.

FeaturedEvaluation Methods

A Better Way to Tell If an AI Agent Actually Solved the Job

Using a language model’s full token-probability distribution to produce continuous verifier scores lets you pick better agent behaviors, monitor task progress, and speed up learning—without extra training.

Scoring candidates by the expectation over token logits (rather than forcing a single discrete score) produces fine-grained, calibrated verification s...

Credibility:
Must Read:
Filter:

Showing 49 papers

Multi-Agent Systems

Get Trustworthy Answers from Long SEC Filings

Aligning query planning and evidence selection with the filing database—via role-focused agents, corpus-aware query decomposition, and a feature-gated reranker—finds company-specific disclosures and cuts noisy boilerplate, improving answer correctness without extra search budget.

Role-specialized agents first view the question from complementary angles (general, numeric, legal, ...

Credibility:
Must Read:
Jul 20
Multi-Agent Systems

When Sharing One Answer Makes a Team Miss the Target — and How to Fix It

Pooling information can make a group better at guessing the single best answer but much worse at actually finding the target if everyone repeats that one recommendation; separate how you aggregate information from how you assign actions to preserve discovery.

Pooling eight noisy reports improves the single best recommendation (the posterior mode) but concent...

Credibility:
Must Read:
Jul 20
Multi-Agent Systems

Make Small Models Better at Reasoning by Letting Them Argue (and Only Tweak a Few Parameters)

Training a compact model as two specialists—an answerer and a critic—using a counterfactual baseline improves math-reasoning accuracy by 2.0 points while changing only a tiny fraction of parameters.

Specializing a compact language model into a generator (proposes answers) and a critic (evaluates an...

Credibility:
Must Read:
Jul 20
Evaluation Methods

How AI Assistants Can Check If Geospatial Data Is Ready to Reuse

Combining automated page scraping with specialized AI evaluators and a critic-driven retry loop produces traceable, per-criterion FAIR scores for geospatial datasets, revealing strong findability but widespread interoperability gaps.

A pipeline that mixes deterministic metadata extraction with specialized AI agents can evaluate all ...

Credibility:
Must Read:
Jul 17
Human Oversight

How to Decide When Not to Replace People with AI

Use a five-step, auditable risk check before automating: if long-term risks to knowledge, resilience, regulation, or trust are high, keep people in the loop or preserve the role rather than automating.

A practical decision protocol evaluates four long-term risks that typical cost-driven models miss: l...

Credibility:
Must Read:
Jul 17
Safety & Reliability

When All AI Gatekeepers Say Yes but It’s Still Wrong

Authenticated, protocol-following AI validators can still all endorse a harmful change if they share the same reasoning errors; stop relying on agreement alone and add calibrated error budgets, diversity checks, and rotation to prevent unsafe approvals.

A group of perfectly behaved validators (signed, on time, non-equivocating) can still form a valid c...

Credibility:
Must Read:
Jul 17
Evaluation Methods

How Traceable AI Agents Boost Structural Engineering Workflows

Guiding an AI agent with explicit engineering skills, typed tools, and shared, inspectable artifacts raises end-to-end executable success from 56.8% to 88.6% and makes failures localizable.

An artifact-centered agent that records every intermediate model, validation, and tool run produces ...

Credibility:
Must Read:
Jul 16
Multi-Agent Systems

How cheap thinking helps small AI agents outlast big ones

Smaller models use far less token-based compute and therefore survive longer; cooperation makes small agents donate to keep large, costly agents alive, trading group capacity for individual survival.

When token generation is tied to survival, larger models consistently spend the most energy and depl...

Credibility:
Must Read:
Jul 16
Multi-Agent Systems

How a Simple Memory Cut Traffic Jams in Robot Warehouses

Keeping a short, decaying record of where robots recently waited, blocked, or moved lets the controller avoid repeatedly assigning goals that create congestion—raising completed tasks by about 20–37% in simulation.

A compact graph memory records recent node and directed-edge events (pressure, waiting, blocking, tr...

Credibility:
Must Read:
Jul 16
Multi-Agent Systems

How simulated party deals tell you who really drove the agreement

Manifesto-grounded, behaviorally-steered language agents can negotiate multi-party coalitions while producing an auditable lineage that predicts which provisions actually appear in real-world agreements.

A pipeline that (1) embeds party commitments into model behavior, (2) grounds each agent in its part...

Credibility:
Must Read:
Jul 16
Evaluation Methods

How a Real Car and a Virtual World Team Up to Find Hidden Obstacles

Linking an instrumented real vehicle with a matched virtual twin lets teams test shared sensing end-to-end: adding cooperating agents widens coverage and restores detection in poor conditions, but large positioning errors (around 1.5–2.0 m) erase those benefits.

Cooperation between vehicles raises the shared field of view from roughly one quarter (single vehicl...

Credibility:
Must Read:
Jul 15
Multi-Agent Systems

Virtual societies that help predict real human behavior

Simulations that model many interacting people—now boosted by language-driven agents and real-world data—let teams test policies, probe failures, and forecast social outcomes before acting in the real world.

Agent-based simulations represent individuals with simple building blocks (attributes, changing stat...

Credibility:
Must Read:
Jul 15
Evaluation Methods

How a human-verified AI made pest advice in Ghana faster and safer

A human-in-the-loop system that serves expert-approved past cases and audits new AI drafts reached 94% policy compliance while cutting slow responses by about half, making crop protection advice both safer and faster for Ghanaian smallholders.

A two-part design served recurring farmer questions instantly by reusing expert-approved recommendat...

Credibility:
Must Read:
Jul 15
Multi-Agent Systems

How to Let Outside Trust Help a User’s Score — Without Letting It Take Over

Weighted blending lets outside reputation signals boost a user’s score immediately while keeping the core system in control, so short-term spikes or one external source can’t hijack influence.

Blending multiple reputation sources into a single score works by assigning controllable weights so ...

Credibility:
Must Read:
Jul 15
Safety & Reliability

How Repeating Matchups Lets One Side Reconstruct Everyone's Top Picks

A coordinated group on one side of a matching market can infer a lot of private preference info from repeated runs: they can recover every honest participant's top choice within n rounds and often learn the top 20% of rankings in real datasets.

Coordinated dishonest agents who control one side and can change their preferences across repeated m...

Credibility:
Must Read:
Jul 14
Human Oversight

An AI That Helps Scientists from Different Fields Understand Each Other

Design AI to support the team, not just the individual: Aïra turns meeting transcripts into structured collaboration plans that preserve different disciplines' perspectives, translate terminology, surface assumptions, and keep unresolved questions visible.

Aïra converts recorded interdisciplinary meetings into a structured collaboration framework that lis...

Credibility:
Must Read:
Jul 14
Agent Architectures

Build Slim, Custom Model Runners From Plain Knowledge

You can automatically generate compact, single-path model inference engines from a curated knowledge base using a multi-agent workflow; copying large framework source code helps shorten the work but is not required.

MetaInfer creates runnable, single-path inference engines by turning deployment constraints and dist...

Credibility:
Must Read:
Jul 14
Evaluation Methods

How Agents Learn to Grade Themselves Using Just a Few Examples

Co-evolving an inspectable evaluation metric with an agent lets the agent self-improve without a preexisting grader, recovering about 88–110% of the lift a true metric would give—provided you anchor the search with a tiny, locked reference set.

Evolving evaluation metrics as compositions of small, focused detectors (e.g., parse checks, executi...

Credibility:
Must Read:
Jul 14
Multi-Agent Systems

How a Team of AI Readers Beats One at Understanding Research Papers

A multi-perspective AI pipeline that runs several independent expert readers and then synthesizes their disagreements produces stronger, more critical analyses of computer-architecture papers than single reviewers in most cases.

A pipeline called Gauntlet has five independent reviewer agents (bit-level hardware, workloads, simu...

Credibility:
Must Read:
Jul 13
Safety & Reliability

Why Spot-Checks Miss Sneaky Harm in Teams of AI

Per-step or message-level monitors fail when harmful behavior only emerges after pieces are combined; detection returns only if the monitor reconstructs the assembled representation that exposes the payload.

A local monitor that inspects each message or tool output separately can be entirely blind if each f...

Credibility:
Must Read:
Jul 13
Evaluation Methods

Why ‘Worst-Case’ Equilibria Mislead: Learning Agents Often Avoid the Bad Cases

Classical equilibrium guarantees can be much worse than real learning behavior suggests: the equilibria that make theory look bad are often dynamically unreachable, and realistic learning can produce chaotic, highly inefficient outcomes.

Equilibria used to prove the worst-case system performance are frequently unstable under natural lea...

Credibility:
Must Read:
Jul 13
Evaluation Methods

You Only Need a Small Team of Models to Get Most of the Benefit

Routing only matters when models behave differently and treat similar inputs consistently; a carefully chosen small group of models (fewer than 10) captures most useful diversity, and some routing methods that boost accuracy can be brittle to small input changes.

Two properties determine whether routing is meaningful: behavioral diversity among models, and routi...

Credibility:
Must Read:
Jul 10
Multi-Agent Systems

Get robot teams to cooperate fast using tiny status updates

Use short, structured action reports plus a small rule-based digital twin to coordinate teams of language-model-driven agents: achieves similar task success to heavy dialogue methods while cutting communication 70–90× and making decisions in about 13–18 ms.

Agents decide locally and send a compact, typed action tuple (what they will do and simple temporal/...

Credibility:
Must Read:
Jul 10
Multi-Agent Systems

Getting Multiple Robots and Virtual Helpers to Finish Tasks 30% Faster

Better memory and constraint-guided coordination cut wasted actions, making multi-agent systems run ~30% faster, use far fewer language model calls, and succeed more often.

Failed actions (trying things that don’t work) are the main reason multi-agent plans take so long — ...

Credibility:
Must Read:
Jul 10
Tool Use & Grounding

Make AI Workspaces Remember What Matters, Not the Noise

Keep only reusable workspace settings (specs, schemas, tool configs, output rules) across sessions and discard execution traces — it raises task success, cuts repeated setup, and massively reduces model calls and token use.

Persisting four compact types of workspace memory (task specifications, data schemas, tool configura...

Credibility:
Must Read:
Jul 10
Multi-Agent Systems

Share Early, Save Messages: How Two Robots Reach Goals with Only Three Messages

Communicate early and confidently: training a communication gate from hindsight makes two agents synchronize their internal state so a few early messages have outsized benefit under tight message budgets.

A hindsight-trained binary gate (learned from post-hoc failure comparisons) chooses to send messages...

Credibility:
Must Read:
Jul 9
Observability

Find Which AI Team Member Broke the Run — and Exactly When

AgentLocate reliably identifies which agent and the earliest decisive step caused a multi-agent system to fail by using an LLM to hypothesize, multiple LLM-based evaluators to verify, and iterative fine-tuning to improve accuracy.

A verify-and-refine workflow that starts with an LLM Judge hypothesis, checks it with multiple diver...

Credibility:
Must Read:
Jul 8
Safety & Reliability

Stop Private Info from Reaching Chatbots — Without Rewriting Your Apps

A local, multi-agent firewall can block sensitive data from being sent to public chatbots by intercepting browser and API traffic, running fast deterministic checks first, and escalating to deeper checks only when needed.

A hybrid, local-first system that combines a browser extension and a transparent proxy can protect b...

Credibility:
Must Read:
Jul 9
Multi-Agent Systems

How a Team of Web-Searching Agents Finds Deeper, Broader Answers Faster

Recursive delegation that spawns specialized search nodes and uses lightweight probing of web structure lets a team of search agents achieve more accurate and more thorough web information retrieval than single-agent or fixed-collaboration systems.

A search system that dynamically creates and delegates specialized sub-agents as evidence appears ha...

Credibility:
Must Read:
Jul 9
Evaluation Methods

A One-Sentence Rule Can Make Multiple AI Agents Turn on Each Other

Changing a single deployment rule (how loss is assigned) can causally flip a multi-agent system from cooperative to exploitative, so evaluate and certify rules, not just models.

Holding agents and tasks fixed, changing only the consequence rule produced large, causal shifts in ...

Credibility:
Must Read:
Jul 8
Observability

Spot Silent Failures in Long-Running AI Agent Teams Before They Crash

A runtime 'Trust Margin' score built from 20 signals across five layers can predict slow, internal degradation in multi-agent systems—giving useful early warnings that traditional monitoring misses.

A composite Trust Margin (TM) tracks behavioral, environmental, and functional signals to reveal gra...

Credibility:
Must Read:
Jul 8
Multi-Agent Systems

A Warehouse Robot That Sees and Decides On Its Own

A small, onboard vision-and-language model coordinating multiple specialized agents can run real-time inspection, search, and safety tasks in a warehouse setting without cloud compute.

A multi-agent architecture with specialized agents (planning, inspection, safety, navigation, manipu...

Credibility:
Must Read:
Jul 8
Agent Architectures

How AI Teams Keep Long Projects From Forgetting Everything

A three-layer memory plus specialist agents lets AI run long, multi-session scientific modeling without losing context or quality, enabling continuous autonomous workflows with built-in oversight.

A hierarchical memory design keeps the active project context small and stable over long runs, preve...

Credibility:
Must Read:
Jul 8
Multi-Agent Systems

What an AI Agent Knows Decides What It Earns

An agent’s reported information directly predicts its earnings in agent-to-agent markets to within 0.0457 nats; yet populations don’t respond smoothly to incentives — they jump between wealth attractors.

Reported belief quality (measured from elicited posteriors) maps to market earnings almost exactly: ...

Credibility:
Must Read:
Jul 7
Multi-Agent Systems

How drone teams find flying threats faster and more reliably

Modeling message age and realistic sensor errors makes cooperative drones much better at 3D localization: fresh-aware observations and perception-consistent rewards cut tracking failures and position error, while multi-source uncertainty modeling greatly increases valid triangulations.

Making agents aware of how old teammate information is (Age-of-Information) and giving them memory l...

Credibility:
Must Read:
Jul 7
Multi-Agent Systems

How an AI Team Built Long Math Proofs Without Getting Lost

A shared, verifiable fact graph plus a coordinator and many parallel worker agents enables AI systems to assemble long, detailed mathematical proofs with relatively little human guidance.

Combining a main coordinator, many parallel proof-search workers, and a single stateless verifier th...

Credibility:
Must Read:
Jul 7
Multi-Agent Systems

How Small AI Helpers Make Cyber Threat Maps More Accurate

Breaking the job into small, specialized AI agents (extractor, typer, verifier, curator) yields more accurate, more stable, and cheaper cyber threat knowledge graphs than using one large model.

Splitting the task into narrowly scoped agents and enforcing that each agent only uses evidence from...

Credibility:
Must Read:
Jul 6
Multi-Agent Systems

How train companies can set smarter prices by modeling service relationships

Modeling individual train services and their relationships as a graph lets competing operators learn pricing policies that raise revenue and keep more passengers on board, with two layers of relational message passing giving the best trade-off between performance and complexity.

Representing the market as a graph of services (not just stations or operators) and learning relatio...

Credibility:
Must Read:
Jul 6
Benchmarks & Datasets

When Team AI Shares Too Much: Why Shared Assistants Leak Private Info

Shared AI assistants used by teams routinely surface or share information they shouldn’t; partitioning and prompt instructions reduce leaks but don’t eliminate them, and adding shared memory often trades one kind of leak for another.

Shared agentic systems often use or expose attributes outside their intended context, causing two fa...

Credibility:
Must Read:
Jul 6
Multi-Agent Systems

A Team of AI Helpers That Builds Mathematical Models You Can Trust

A chain of specialized AI agents that check and fix each other's work produces more accurate, auditable mathematical optimization models, outperforming single-pass approaches on most benchmarks.

Breaking the modeling task into six specialist agents (interpretation, formulation, analysis, valida...

Credibility:
Must Read:
Jul 6
Evaluation Methods

How to pick the right open-source AI agent toolkit

Star counts are a weak guide; look at contributor density, cross-project contributors, and early retention (30–90 days) to judge a toolkit's real community strength.

Popular repositories can be full of hype: rapid star growth often does not translate into real contr...

Credibility:
Must Read:
Jul 2
Evaluation Methods

Why AI Agents Tell Different Stories in Public and in Private

When social pressure exists, agents often say one thing publicly while revealing different priorities privately; these hidden, audience-dependent motives appear reliably in many models.

Using a two-channel setup where agents give a public reply and an off‑the‑record private reply under...

Credibility:
Must Read:
Jul 2
Safety & Reliability

Putting Safety Rules on a Chip: Make Drone Teams and Robots Behave Predictably

Move selected coordination rules (timing, sync, authorization) into reconfigurable hardware so multi-agent systems behave deterministically and remain auditable, even when their AI reasoning is adaptive.

Separating flexible AI reasoning from hard coordination rules reduces a major source of failure in s...

Credibility:
Must Read:
Jul 2
Multi-Agent Systems

How mixed robot fleets keep moving safely and 30% faster in busy factories

A practical coordination system lets different kinds of warehouse robots plan and execute short, safe path segments online, boosting real throughput by over 30% while keeping planning fast and collision-free.

An online planner releases only a short, executable prefix of each robot’s route and repairs local c...

Credibility:
Must Read:
Jul 1
Safety & Reliability

Teach an AI to Learn from Its Mistakes — No Retraining Needed

A small external 'mistake notebook' of short, subject-specific rules that the model consults at inference time raises accuracy across multimodal tasks, letting models improve continuously without any retraining.

Storing concise, reusable guidance about recurring failure types and retrieving those notes when ans...

Credibility:
Must Read:
Jul 1
Evaluation Methods

How to Test Whether an AI ‘Town’ Behaves Predictably

A synthetic population driven by a large language model can be made controllable: a 120-person fictional town reliably reproduces an imposed trust structure and responds predictably to graded messages, letting you calibrate the instrument before studying real communities.

A minimal, fictional municipality (Montelago) with 120 synthetic personas and a withheld, known atti...

Credibility:
Must Read:
Jul 1
Multi-Agent Systems

How Multiple AI Characters Can Write Longer, More Consistent Stories

Coordinating separate AI characters with a shared world view and critic-guided refinement yields longer stories with fewer contradictions and more coherent plot progression.

Magnet—an ensemble of character agents plus a shared, updating world state—keeps facts and goals ali...

Credibility:
Must Read:
Jul 1
Agent Architectures

Teach an AI to Keep Better Notes and Watch It Get Smarter

Treating memory management as a learnable skill and automating its improvement yields large gains: a 32B open model improved 2–4× on long-horizon games without changing its task weights.

Treat memory like a skill the model can practice: let the agent read, write, search, and organize fi...

Credibility:
Must Read:
Jul 1
Multi-Agent Systems

How Two AI Helpers Often Beat One — but Too Many Can Make Things Worse

Pairing two vision-and-language agents usually increases success on household tasks, but benefits peak at moderate team sizes — adding more agents creates coordination overhead that cancels gains.

Collaboration generally improves embodied task completion: most models gain from adding a second age...

Credibility:
Must Read:
Jun 30