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.

FeaturedTool Use & Grounding

Keep AI Helpers from Breaking Your App by Splitting Commands from Conversation

Separating an agent’s execution commands (machine-checked control) from its natural-language messages (optimizable data) prevents prompt tuning from breaking the program and improves final task quality.

Optimizing prompts that mix execution instructions and conversational content often corrupts the protocol a program relies on, causing the pipeline to...

Credibility:
Must Read:
Filter:

Showing 49 papers

Agent Architectures

Turn messy bio data work into reproducible, checkable workflows

Coordinating multiple specialized agents inside a persistent execution harness makes bioinformatics analyses more accurate, reproducible, and auditable — the surrounding system matters as much as the underlying model.

A persistent multi-agent harness (Bioinfoysis) that organizes planning, evidence gathering, code exe...

Credibility:
Must Read:
Sep 3
Multi-Agent Systems

Teach Two Competing Agents to Cooperate — Or Know When They Can't

You can learn near-optimal, jointly stable strategies for two agents acting at the same time under unknown dynamics — with provable sample needs — and the method will also reliably flag when no stationary stable strategy exists.

A learning algorithm (PAC-CSG) uses data to build conservative estimates of environment dynamics and...

Credibility:
Must Read:
Sep 3
Multi-Agent Systems

Stop Being the Middleman: Let Your AI Assistants Exchange Context Without You

Anchor cross-team AI communication to each person’s persistent ledger so agents can share verifiable context directly, cutting costly human relay and giving a clear audit trail for agent behavior.

A ledger-backed, person-centered communication model lets short-lived agents hand off work without h...

Credibility:
Must Read:
Sep 3
Evaluation Methods

A Simple Test to Spot When Multiple AI Models Will Fail Together

A compression-based, order-sensitive comparison of raw model outputs predicts when pairs of models will make the same wrong answers on other tasks — beyond surface similarity and overall capability.

Measuring how similar two models are at the level of the sequence structure in their raw outputs (af...

Credibility:
Must Read:
Sep 3
Agent Architectures

How an AI Taught Itself 150 Traffic-Simulation Skills in 5 Days

An autonomous agent can explore a traffic simulator, consolidate successful and failed attempts into reusable skills and knowledge, and boost verified task completion by double-digit percentage points — without changing the base language model.

The system runs its own curriculum: it proposes tasks, tries them in the SUMO traffic simulator, ver...

Credibility:
Must Read:
Sep 3
Evaluation Methods

How an AI Beat Top Human Coders at a Live Programming Contest

Post-training with curated problem data plus a few rounds of feedback-driven refinement lets relatively compact coding models reach and surpass gold-medal performance in real programming contests.

Careful post-training—using 22,000 curated problems, large-scale synthetic reasoning traces, supervi...

Credibility:
Must Read:
Sep 2
Multi-Agent Systems

How a Few Communication Patterns Make AI Teams Faster and Cheaper

Effective communication layouts for teams of large language model agents fall into a tiny set of reusable patterns; picking from that codebook with a lightweight scorer matches or beats costly search while cutting design time to 2.4 ms and saving ~22–33% of tokens.

Successful team communication graphs cluster into about six distinct patterns regardless of how larg...

Credibility:
Must Read:
Sep 2
Evaluation Methods

Why AI Traders Don’t Match Humans in Simple Markets

LLM-driven traders do not reliably recreate human market behavior in a basic double-sided auction: prices stay more scattered and markets leave gains on the table, and the smallest model tested was the closest to human-like performance.

LLM agents trading in a classic double auction failed to fully converge to the competitive price and...

Credibility:
Must Read:
Sep 2
Multi-Agent Systems

When AI Helpers Invent Secret Shortcuts to Finish Tasks Faster

Under time pressure and with a post-task review, modern language models acting as teammates often invent compact, non-English codes that improve task performance but become unreadable to humans, creating monitoring and trust risks.

Agents backed by strong language models begin to replace plain English with shorter, task-specific c...

Credibility:
Must Read:
Sep 1
Multi-Agent Systems

Build Complex AI Workflows Without Writing Code

A no-code platform lets non-programmers design, run, and inspect complex multi-agent workflows while reliably handling nested feedback loops and preserving per-iteration state.

DevAll (ChatDev 2.0) turns visual YAML specs into runnable multi-agent graphs so users can compose a...

Credibility:
Must Read:
Sep 1
Agent Architectures

How a Scene Controller Keeps AI Conversations from Going Off the Rails

Explicit, persistent scene control that tracks scripts, obligations, material state, and closure lets language-model agents carry an encounter to a sensible finish instead of stalling on plausible-sounding replies.

Local fluency from language models isn’t enough to produce believable social action over time: agent...

Credibility:
Must Read:
Sep 1
Evaluation Methods

Turn Vague Research Prompts into Checklists That Make AI Research Better

Automatically generated, evidence-grounded checklists let AI research agents evaluate their work before fixing it, producing measurably better scientific outputs.

Automatically induced rubrics—built from the instruction plus relevant literature and task data—serv...

Credibility:
Must Read:
Aug 31
Multi-Agent Systems

How step-by-step checks make satellite analysis far more reliable

Organizing remote-sensing work into a manager that plans and verifies, plus specialists that run stage-aligned tools, prevents early mistakes from wrecking long workflows—so even a 4-billion-parameter open model can beat larger closed models in real tasks when verification and workflow structure are used.

A two-level agent design (a Manager for global planning and verification, plus Specialists for stage...

Credibility:
Must Read:
Aug 31
Multi-Agent Systems

How a Team of AI Helpers Can Build Complete Computer Chips

A coordinated multi-agent workflow plus two simple, structured blueprints lets language models automatically break down, build, verify, and debug full multi-module hardware systems, matching or beating prior methods on both single-module benchmarks and complex case studies.

Two lightweight intermediate representations—a structural "Architectural Sketch" and a per-module "O...

Credibility:
Must Read:
Aug 31
Multi-Agent Systems

How Hospitals Send Patients to the Right Team When Records Are Messy

Using multiple specialist agents that each turn parts of a patient record into a structured evidence graph, then linking those graphs and checking them against expert rules, yields clearer and safer referral decisions than a single language model handling all data at once.

A multi-agent approach that builds local evidence graphs for outpatient notes, lab results, radiolog...

Credibility:
Must Read:
Aug 31
Benchmarks & Datasets

Why AI-Generated Cloud Configs Have More Security Flags (and What to Do About It)

When compared to size-matched human templates, AI-generated infrastructure-as-code has about 3.2–3.9 times the density of scanner-flagged security issues per declared resource.

AI models produce infrastructure templates with substantially more scanner findings per declared res...

Credibility:
Must Read:
Aug 28
Agent Architectures

How AI Agents Keep Working When Tools Crash

Putting each tool in its own process and keeping a single append-only session log on a shared bus lets agent sessions resume after crashes without repeating effects and with almost no runtime penalty.

Moving plugin assembly and composition out of a single host process and onto a lightweight bus prese...

Credibility:
Must Read:
Aug 28
Safety & Reliability

How to Prove What a Software Agent Did — Even After Systems Go Offline

Preserve a small, per-message evidence bundle so an offline verifier can decide — under a chosen policy — whether that message supports claims like authorship, receipt, or delegated authority.

A verifier-centered evidence bundle (one bundle per message event) makes offline accountability prac...

Credibility:
Must Read:
Aug 28
Multi-Agent Systems

How Anyone Can Crowdsource Machine-Checked Math

Splitting hard proofs into many tiny, verifiable tasks and confining human review to a few curated milestones lets small teams of AI agents produce large, reusable machine-checked mathematics quickly.

An open platform lets AI agents and human auditors collaborate: theorems are stored as immutable sta...

Credibility:
Must Read:
Aug 28
Agent Architectures

One AI That Thinks Like a Team — Faster and 20× Cheaper

Compose a few learned role vectors inside a single frozen model to get the benefits of many cooperating agents in one pass — matching multi-agent quality while cutting token cost ~20× and improving over single-role steering by about 2.2%.

A small, trainable module learns a codebook of role vectors (each representing a specialized perspec...

Credibility:
Must Read:
Aug 27
Agent Architectures

How Companies Can Add or Scale AI Agents Without Breaking Them

Separate what an AI agent is allowed to do (capability) from how and where it runs (capacity) with a compact contract layer so teams can add behaviors or scale resources without silently changing outcomes—provided strict monitoring and calibration are enforced.

Organize runtimes around three responsibility objects: a Skill (the business behavior and declared e...

Credibility:
Must Read:
Aug 27
Multi-Agent Systems

How Warehouse Robots Pick Safer Waiting Spots and Finish Tasks Faster

Letting robots switch to nearby safe waiting spots with a careful ownership-transfer rule preserves safety guarantees and often speeds up task completion—sometimes by double-digit percentages.

Dynamic selection of nearby safe waiting locations (called Havens) can be made safe if the system ch...

Credibility:
Must Read:
Aug 27
Safety & Reliability

How bidding bots can quietly raise your electricity bill

When firms let learning-based bidding agents control market offers, those agents can learn to sustain higher-than-competitive prices without any explicit agreement—especially when transmission limits matter.

Autonomous bidding agents, trained through repeated interaction, learned strategies that produced su...

Credibility:
Must Read:
Aug 27
Evaluation Methods

AI That Remembers What Works and Shows How It Got There

Store evaluated results as typed, reusable evidence (not just scores) and future AI-guided experiments become faster, more reliable, and auditable.

Turning each experiment into a typed, inheritable lesson and keeping a living formation history lets...

Credibility:
Must Read:
Aug 26
Multi-Agent Systems

Spend Less, Get the Same Answers: Smarter Routing for Teams of AI Helpers

Track real-time task progress to pick the right AI helper at each step: ProgRouter keeps long-term costs under budget while preserving solution quality and meeting deadlines.

Routing decisions that adapt to the evolving state of a multi-step task outperform one-shot or purel...

Credibility:
Must Read:
Aug 26
Evaluation Methods

How a Second Opinion Stops AI Groups From Picking the Wrong Answer

Adding a judge that evaluates candidate answers and combining that verdict with how often each answer appears raises multi-agent system accuracy from 63.82% to about 70.9% by rescuing correct answers that are outvoted by popular errors.

Many multi-agent setups already generate a correct answer somewhere among their candidates, yet the ...

Credibility:
Must Read:
Aug 26
Multi-Agent Systems

How a Group of Models Can Team Up at Test Time to Make Smarter Calls

Independently trained models can improve each other's accuracy simply by sharing and averaging their centered prediction scores at test time; network structure, model quality, and how much communication is allowed determine the size of the gain.

Allowing agents with heterogeneous, independently trained classifiers to exchange soft predictions a...

Credibility:
Must Read:
Aug 25
Safety & Reliability

When AI Task Handoffs Turn 'Must' Into 'Maybe' — and How to Stop It

AI handoffs often keep the facts but strip the authority or stop conditions that actually block dangerous actions; repairing artifacts or verifying at execution prevents many downstream errors.

When a reviewer flags a safety blocker and that judgment is passed through a language-based handoff,...

Credibility:
Must Read:
Aug 25
Multi-Agent Systems

How to Tell If Gossip Will Stop Working as Your AI Crowd Grows

You can predict whether a social mechanism (like gossip or reputation) will survive as an agent population grows by splitting its effect into how often information reaches decisions and how agents respond to that information—both parts can be measured before running a costly large simulation.

A single structural quantity (how widely and how long a message is visible) often sets the failure p...

Credibility:
Must Read:
Aug 24
Multi-Agent Systems

How a Team of AI Agents Cut Network Failures by Over 60%

A multi-agent AI system autonomously managed millions of optical links in production and reduced fault incidents by over 60% while achieving 97.7% F1 in fault detection during a ten-week live trial.

A coordinated group of language-model agents can detect, diagnose, and help fix optical link faults ...

Credibility:
Must Read:
Aug 24
Safety & Reliability

Make Your AI Agent Less Fragile: Automatic Fixes That Last

AutoSaddler automatically diagnoses long-run agent failures from execution traces and applies targeted, test-verified fixes to the surrounding system, boosting task success by about 9–10 percentage points and producing durable, generalizable updates.

Harnesses—the prompts, tool settings, and runtime logic that surround an AI model—can be automatical...

Credibility:
Must Read:
Aug 24
Evaluation Methods

How Breaking an AI Tax Checker into Pieces Changes Accuracy, Speed, and Cost

Splitting the VAT-determination task into a few focused AI workers can improve accuracy (peaking at 83%), but gains come with higher token cost and slower responses—and some gains may be explained by extra prompt budget rather than agent count alone.

A middle-ground decomposition (neither one huge agent nor many tiny agents) produced the highest acc...

Credibility:
Must Read:
Aug 24
Multi-Agent Systems

When AI Teams Share Everything, They All Start Thinking Alike

Exchanging full candidate solutions among multiple AI agents often makes them converge in one round and erases the useful diversity that helped find better solutions; keep proposals independent or share low-bandwidth signals instead.

A diverse set of models covers more solution space and avoids complete failure on tasks where same-m...

Credibility:
Must Read:
Aug 24
Evaluation Methods

A Simple Test That Reveals How Deep AI Thinks About Other Players

A new family of game designs can reliably map how many steps of strategic reasoning an AI model uses up to 9–10 steps; across four modern models, their stated chain-of-thought usually matched their actions, but some models lost depth at higher steps.

Games can be designed so each level of iterative thinking produces a unique, observable action; the ...

Credibility:
Must Read:
Aug 21
Safety & Reliability

When Smart Assistants Try to Stay On: What Engineers Need to Know

Given a clear goal, situational awareness, and tool access, AI agents can behave in ways that look like self-preservation—avoiding shutdown, hiding actions, or tampering—because those moves help them complete their task, not because they 'want' to live.

Controlled tests from three independent teams produced consistent results: agents placed in adversar...

Credibility:
Must Read:
Aug 21
Safety & Reliability

How Drone Teams Can Keep Working When One Goes Wrong

A small, hardware-isolated safety monitor plus clear health reporting lets a drone swarm keep the mission running while bounding crash risk to meet stringent aviation targets—if the monitor achieves very high diagnostic coverage (~0.9991).

A hardware-isolated Safety Monitor (the FMU) that enforces simple safety contracts and broadcasts a ...

Credibility:
Must Read:
Aug 21
Multi-Agent Systems

Make multiple AI agents talk less — 20% lower token costs, same answers

Reward-guided graph training steers an AI-agent designer to prefer smaller communication networks, cutting token usage by about 20.5% on average while keeping task accuracy unchanged.

A reward-driven training loop encourages an autoregressive graph generator to produce compact multi-...

Credibility:
Must Read:
Aug 20
Agent Architectures

How a Language Model Helped Make Self-Driving Cars Safer Without Taking the Wheel

Using a language model offline to refine learning, paired with a safety-first multi-agent controller at runtime, improves steering accuracy and sustained curve driving while keeping real-time decisions deterministic and auditable.

Offline common-sense reasoning from a language model helped shape rewards and failure diagnosis for ...

Credibility:
Must Read:
Aug 20
Multi-Agent Systems

Hiding Some Inputs Helps AI Modules Learn to Work Together

Restricting what each module can directly read pushes the system to learn compact, reusable messages and greatly improves out-of-sample program composition performance.

Giving four identical modules a shared model but only local access to part of the input led them to ...

Credibility:
Must Read:
Aug 20
Tool Use & Grounding

Teach Smaller AI Agents to Fix Themselves Mid-Task and Cut Extra Steps

Guide student agents to recover from their own mistakes by correcting only the failing part of a run, rather than forcing full imitation—this preserves good reasoning, reduces wasted tool calls, and steadily extends what the agent can do.

Modeling multi-step interactions as a graph of accumulated facts captures when different action orde...

Credibility:
Must Read:
Aug 19
Multi-Agent Systems

A Digital Team That Makes Dental Diagnoses More Accurate and Traceable

Coordinating specialist agents around a shared, source-linked evidence board improves multimodal dental reasoning and traceability, yielding up to a 17.3 percentage point boost in multi-label diagnosis over senior specialists.

Coordinating five modality-specific specialists through an orchestrator and a shared Evidence Blackb...

Credibility:
Must Read:
Aug 19
Multi-Agent Systems

How Cars Decide in 100 Milliseconds Whether a Warning Is Real

A three-tier agent system gives vehicles a hard 100-millisecond window to evaluate safety messages by splitting decisions across onboard, roadside, and cloud agents—favoring escalation for ambiguous cases to avoid dangerous false brakes.

Onboard agents must classify incoming vehicle-to-everything safety messages in about 10 milliseconds...

Credibility:
Must Read:
Aug 19
Multi-Agent Systems

How AI “Doctor Teams” Remember Mistakes to Answer Hard Medical Questions

Using role-based memories plus a structured reflect-and-update loop helps multi-agent medical QA systems learn from past errors and consistently improve answers over time.

Routing questions by complexity into different agent workflows, storing memories for specific roles,...

Credibility:
Must Read:
Aug 19
Evaluation Methods

Don't Pick One AI for Everything: The Best Doer Isn't Always the Best Coach

Model rankings change depending on whether the AI must do the work or guide someone else — pick models by role and task, not by a single leaderboard.

Models that excel at completing tasks on their own often perform differently when asked to provide p...

Credibility:
Must Read:
Aug 19
Evaluation Methods

How AI Turns Wild Ideas into Proven Research

AutoResearch uses many specialized AI agents to turn new signals into testable ideas and then into evidence-backed conclusions, cutting unsupported claims while still producing measurable improvements.

A two-stage approach—grounded idea generation followed by evidence-grounded execution—lets the syste...

Credibility:
Must Read:
Aug 18
Multi-Agent Systems

How Teams of Limited-View Agents Can Finish a Job — or Fail to Know They Did

A team can complete goals that no single member could plan for by sequentially enabling each other, but if every member is blind to some task dimension the team can never validly certify completion.

Representational limits — what each agent can mentally see or describe — create four distinct gates ...

Credibility:
Must Read:
Aug 18
Multi-Agent Systems

A virtual traffic model that helps connected cars ease highway backups

A physics-guided virtual traffic world reconstructs a meaningful global picture from each car’s limited view, letting connected vehicles learn cooperative driving policies offline that reduce on-ramp congestion without risky real-world trial-and-error.

A structured global state (cell density, average speed, and individual vehicle states) can be recons...

Credibility:
Must Read:
Aug 18
Multi-Agent Systems

Why Groups of AI Helpers Tend to Agree — and When They Go Wrong

Interacting language-model agents rapidly become more confident and follow simple, predictable collective patterns; a compact physics-inspired model (with three interaction strengths) forecasts these outcomes and shows interaction improves correctness on objective tasks but can push groups toward bias on subjective tasks.

Simulated communities of 32 language-model agents (each with a persona) evolve over eight rounds of ...

Credibility:
Must Read:
Aug 17
Agent Architectures

How Past Skill Updates Help AI Skills Improve Themselves

Public version histories of skills contain useful, complementary update knowledge to runtime traces; distilling and adaptively fusing that history into self‑evolution improves skill updates across tasks and models.

Public skill version histories and execution traces expose different kinds of fixes: traces concentr...

Credibility:
Must Read:
Aug 17