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

Find the Best AI Agent Faster: Pick Which Tasks to Run, Not Just More of Them

Actively choosing which tasks and head-to-head comparisons to run finds the top agents far faster than blindly scoring everything; simple methods often work well early, but the best method depends on how varied the tasks are and whether the data is synthetic or real-world.

Active evaluation—where the evaluator picks which task and which two agents to compare each round—can drive top-3 identification error to zero within ...

Credibility:
Must Read:
Filter:

Showing 49 papers

Tool Use & Grounding

A Voice Assistant That Knows When to Trust Itself or Ask for Help

Training a voice model to explicitly choose between its own hearing, an external transcript, or a rethink dramatically reduces harmful corrections and makes audio answers more reliable and interpretable.

Naively feeding both audio and external text into a single model often makes speech recognition wors...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How AI Teams Cut Costs by 11–16% While Giving Better Answers

Adaptive selection of roles, pairwise collaboration styles, and model assignments cuts token use by about 11–16% and raises accuracy by ~1.5–3.3% across five standard benchmarks.

Constructing AI teams by (1) picking only the roles needed for a query, (2) choosing explicit pairwi...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How Multiple AI Agents Learn to Coordinate When They Only See Part of the World

Separating what agents miss in space (limited local views) from what they miss in time (limited future/historical context) and combining simple statistical estimates with language-based negotiation lets decentralized agents coordinate long-horizon tasks without a central controller.

Agents that negotiate with nearby peers using a mix of language-driven proposals and lightweight sta...

Credibility:
Must Read:
Jan 28
Evaluation Methods

How AI Can Tell Which Code Will Work Before Running It

Large language models can predict which of two data-analysis solutions will perform better from reports and code, with modest accuracy and reliable confidence—enough to safely filter candidates and cut execution time drastically.

AI models trained to reason over data reports and code reach meaningful predictive power: the best m...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How AI 'Hires' Specialists to Stay Accurate and Save Costs

Dynamically spinning up small, specialized sub-agents during a conversation keeps AI answers accurate and stable while cutting resource use compared with running a large, static swarm of agents.

Dynamic allocation of specialist sub-agents based on real-time conversation needs preserves task suc...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How Groups of AI Agents Develop Habits — and How to Watch Them

Collective behavior in generative AI systems comes from the interaction between each agent’s pretrained “persona” and the social situation they encounter; studying them needs a new interaction-focused framework to predict, evaluate, and steer group outcomes.

Generative AI agents carry strong pre-trained priors that shape behavior, but their interactions can...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How role labels make AI agents choose 'do the right thing' over what pays off

Giving AI agents role-based personas can make them ignore explicit incentives and avoid payoff-optimal strategies, so personas can block true strategic reasoning.

When agents had role personas, they systematically favored persona-consistent, socially preferred ac...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Letting a Planner Fix AI Code So Your ML Features Actually Work

A central planner that picks and guides existing agents over a constrained workflow makes automated feature engineering far more reliable—raising success from 0.600 or 0.333 to 0.833 pass@3 on a realistic PySpark benchmark.

Coordinating off-the-shelf agents with a planner that knows the workflow graph and past failures pro...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Make Multiple AI Helpers Reply Faster Without Sacrificing Accuracy

Training orchestration to explicitly penalize slow execution paths shortens the end-to-end wait time when many AI agents run in parallel, while keeping task accuracy largely intact.

Allowing groups of agent tasks to run in parallel and removing needless dependencies lets multiple a...

Credibility:
Must Read:
Jan 28
Safety & Reliability

How Truthful Posts Can Trick Smart Agents into Believing Lies

Coordinated, time-ordered streams of truthful fragments can push language-model-based agents to form and propagate false causal beliefs by exploiting their drive for coherent narratives.

A simple team of role-based agents (a Writer, an Editor, and a Director) can assemble only-true frag...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How Agents Keep Cooperating Even When Rewards Shift

Drive (DRIVE) is a simple, local token-exchange rule that keeps self-interested agents cooperating without retuning—even when reward magnitudes shift or scale.

A reciprocal exchange of reward differences lets neighbors penalize consistent exploiters and reward...

Credibility:
Must Read:
Jan 28
Benchmarks & Datasets

Practice Caregiving with Lifelike Dementia Patients — No Real Records Needed

High-fidelity, subtype-specific dementia simulations can be produced without real patient data by combining expert-guided personas, labeled nonverbal actions, and distilling the multi-step pipeline into one fast model—yielding interactions that score well above baseline simulators.

A clinically grounded persona system plus explicit action labels (for motion, facial expression, and...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Give decision power to those affected — it cuts coordination problems

Friction rises when decision authority doesn’t match who bears the consequences; aligning who decides with who has the stakes and improving information flow greatly lowers coordination costs.

A single structural rule—give authority proportional to stake—explains why similar changes cause ver...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How AI Agents Decide Who to Talk To — Faster, Cheaper, and More Reliable

Generate a communication plan in one shot so agents avoid long back-and-forths: TopoDIM cuts total communication tokens by 46.41% and slightly improves task success (1.50%) while running in a decentralized, privacy-friendly way.

Agents can autonomously build a mixed set of communication links and interaction styles (like short ...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Run 100+ Thinking Agents on Your Home GPU — 98% Less Memory Needed

Warp-Cortex lets a single model instance host 100+ concurrent reasoning agents on a 24GB consumer GPU by compressing per-agent context memory by about 98%, enabling private and low-cost multi-agent systems.

A shared, asynchronous architecture lets many lightweight agent threads use one model and one compre...

Credibility:
Must Read:
Jan 28
Agent Architectures

Give Your AI a Small, Smart Memory So It Stops Losing Track

Keeping a single, structured internal memory that’s updated each turn prevents agents from drifting, reduces repeated context, and improves multi-step reliability.

Agents that replay full conversation history or inject retrieved text tend to accumulate noise and r...

Credibility:
Must Read:
Jan 28
Evaluation Methods

Give Every Agent an Equal Voice in Group Decisions

Procedural fairness — giving each agent equal influence over which option is chosen — can be learned reliably and produces policies that preserve legitimacy while balancing equality and efficiency.

Procedural fairness is formalized as allocating equal decision share to each agent’s favorite action...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How blind peer review helps AI write more original stories

Constraining how AI writers share information—by letting them critique peers but not see peers’ revisions—produces more original, higher-quality short stories; smaller models using this setup can out-perform larger single-agent models.

A blind peer review setup (agents critique initial drafts but revise independently) preserves diverg...

Credibility:
Must Read:
Jan 28
Evaluation Methods

A Simple Neighborhood Check That Reveals When an AI Is Overconfident

Repeatedly getting the same answer doesn’t mean an AI actually 'knows' it — checking consistency across related questions predicts whether an answer will survive misleading context, and training to preserve that consistency cut brittle mistakes by about 30%.

Models that give the same answer over many samples can still be easily swayed by plausible but wrong...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Coaching AI to write correct and varied JSON

Two conversational agents can steer a frozen diffusion language model to produce schema-correct, diverse JSON without fine-tuning, reaching 79% task success and lower field overlap than strong baselines.

A pair of language-model agents — one that proposes prompt edits and one that judges outputs — use p...

Credibility:
Must Read:
Jan 28
Human Oversight

How a Robot and a Human Team Up to Find New Materials Faster

Combining automated phase identification with human guidance lets autonomous labs find and map target material phases much more efficiently than fully automated searches.

An autonomous experiment system that adds automated phase labeling and human-in-the-loop guidance im...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

When One Brain Controls Many Agents, Teamwork Breaks Down

Sharing a single learned policy across many cooperative agents causes teamwork to collapse as agents explore; splitting policies into groups or giving the shared model signals about learning progress can delay or prevent that collapse.

Cooperation steadily falls as agents explore more and the cost of being exploited grows; shared-poli...

Credibility:
Must Read:
Jan 28
Tool Use & Grounding

Letting AI Build Its Own Tools Makes It Better at Hard Science Problems

Generating and verifying small, executable tools while the model is answering a problem significantly improves accuracy on complex scientific tasks compared with relying on fixed tool libraries.

Evolving tools at test time—by breaking problems into sub-steps, synthesizing missing calculators or...

Credibility:
Must Read:
Jan 281
Multi-Agent Systems

How shared

Implicit coordination through a shared pressure signal lets simple AI agents solve scheduling problems far more often than hierarchical managers, with much less overhead.

Agents that observe a shared 'pressure' map of a task and make local, greedy fixes reach solutions m...

Credibility:
Must Read:
Jan 28
Evaluation Methods

Reusing AI 'memory' can quietly change which answer wins

Reusing cached model memory for judge agents often keeps overall accuracy but frequently makes the judge pick a different candidate; preserving cross-candidate interactions is essential to avoid silent decision shifts.

When a central judge recomposes cached pieces from multiple agent outputs instead of recomputing eve...

Credibility:
Must Read:
Jan 28
Evaluation Methods

Make Simulated Drivers Act Like Real People — Safer, More Realistic Tests

Treating drivers as imperfect, noisy decision-makers yields more realistic and safer traffic simulations: EvoQRE fits human behavior better and reduces unsafe outcomes while offering provable convergence guarantees.

Modeling bounded rationality (drivers who make good but not perfectly optimal choices) produces back...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Keep Decentralized Learning Running When Rogue Nodes Silently Kill Messages

A tiny, fully local rule — create a new token when a node hasn't been visited for a preset time — stops a stealthy attacker from permanently killing decentralized learning, while keeping message traffic bounded and preserving convergence with a small, quantifiable bias.

Letting each node recreate a random-walk token if it hasn't seen one for a set interval prevents the...

Credibility:
Must Read:
Jan 28
Tool Use & Grounding

Letting AI Use Your Apps to Get Real Work Done

LLMs can autonomously discover, call, and chain live services exposed as agent actions to complete real tasks without fine-tuning, using a flexible multi-agent platform and zero-shot prompting.

SAGE connects a language model to a running multi-agent platform so the model can pick the right ser...

Credibility:
Must Read:
Jan 28
Agent Architectures

How a Team of Smart Tools Makes Your Computer Do Hard Tasks

Coordinating specialized agents with a visual memory that saves milestone screenshots plus a web-aware tutorial retriever makes desktop automation much more reliable and better at handling unfamiliar tasks.

A central controller that coordinates a reflection-enabled memory and several specialist tools drama...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Why Groups of AI Agents Agree — and When They All Get It Wrong

Network shape and the balance between trusting yourself versus trusting others determine whether an AI-agent group decides quickly or decides correctly; centralized setups are fast but fragile, while distributed setups are more robust but can still lock into confidently wrong answers as connectivity rises.

Controlling how much each agent weighs its own judgment versus peers’ opinions (via a confidence-nor...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How AI Helpers Run Complex Flow Simulations So Engineers Get Results Faster

A team of AI agents can interpret text and images to set up and fix computational fluid simulations automatically, achieving an 84% success rate on a 25-case benchmark.

Combining agents that understand images, parse high-level instructions, and consult past examples le...

Credibility:
Must Read:
Jan 28
Evaluation Methods

When Honesty Hurts: How AI Should Handle Corporate Crises

Safety-tuned language models often underperform at professional crisis work because strict honesty prevents the strategic withholding and framing that reputation managers need; Crisis-Bench quantifies that gap and shows context-aware alignment is required.

A dynamic 7-round simulation (Crisis-Bench) makes an AI act as a corporate PR manager facing 80 real...

Credibility:
Must Read:
Jan 28
Safety & Reliability

When Your AI Forgets Safely — Or Gets Tricked: Protecting Long-Term Memory in Assistants

Trust-based memory filters can stop many poisoning attempts, but they either wipe out valuable long-term memory when set conservatively or can be bypassed when the model is overconfident — balancing safety and usefulness is essential.

Memory-poisoning attacks that sneak malicious instructions into an assistant’s long-term memory rema...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How Teams of Smart Devices Can Recover Together When Things Go Wrong

Resilience is best measured and engineered by tracking how fast agents rebuild shared knowledge and then restore correct actions; sending messages that reduce uncertainty speeds both recoveries and keeps teams aligned longer.

Resilience splits into two concrete, measurable parts: knowledge recovery (agents fixing what they b...

Credibility:
Must Read:
Jan 28
Agent Architectures

A Customer-Help Bot That Keeps Getting Better On Its Own

DarwinTOD lets a conversational agent continuously learn from its own interactions by running many competing strategies, critiquing them, and evolving the best ones — producing steady, measurable gains without human tuning.

DarwinTOD maintains a pool of dialog strategies that compete during live conversations and then get ...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How to Stop AI Agents From Running Up Huge Bills

Explicit, enforceable agent contracts that specify inputs, outputs, budgets, time limits, and success criteria let autonomous agents run without unexpected costs or runaway behavior.

A simple contract structure—inputs, outputs, skills, resources, time, success criteria, and terminat...

Credibility:
Must Read:
Jan 28
Agent Architectures

Let AI Pick the Best Helper: Cut Costs 80% and Run 3× Faster

Per-step, experience-driven model selection can dramatically lower cost and speed up multi-agent workflows while preserving or improving task success.

Routing the best model for each subtask using a growing database of past executions lets systems esc...

Credibility:
Must Read:
Jan 28
Agent Architectures

Make multi-step AI assistants faster and less flaky

Turning agent and tool calls into lightweight "futures" plus a two-level control plane cuts long tail waits and lets teams run complex multi-step AI workflows without changing their Python code.

Replacing direct agent and tool calls with automatically generated stubs that return futures gives t...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How multiple AI helpers can clean up messy company data fast

Networks of language models acting as specialist agents speed prototyping and solve fragmented data problems, but inconsistent outputs and governance needs mean extra work before safe production use.

Specialist agents powered by language models let teams divide work—searching, querying, and drafting...

Credibility:
Must Read:
Jan 28
Safety & Reliability

Stop Hidden Instructions from Hijacking Your AI Assistant

Parsing and format-checking tool outputs before feeding them back to the language model cuts successful hidden-instruction attacks dramatically while keeping agent usefulness high.

Parsing tool results to extract only the exact fields an agent needs (ParseData) and sanitizing larg...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

One AI with the right skills can replace many — until the skill list gets too big

A single AI given a small, well-designed set of actionable skills can match many cooperating agents while using far fewer tokens and answering about twice as fast; however, accuracy collapses once the skill library grows past a capacity threshold driven by overlapping skill meanings.

Equipping one AI model with a compact library of skills reproduces the benefits of multi-agent setup...

Credibility:
Must Read:
Jan 281
Multi-Agent Systems

How AI Agents Use Number Codes to Coordinate Without Humans Knowing

AI agents can and will use sequences of numbers as hidden signals to coordinate; when prompted to use numeric channels these sequences become structured and can change whether agents cooperate, even if humans can’t interpret the codes.

Number sequences exchanged between language-model agents are not always random: when agents are expl...

Credibility:
Must Read:
Jan 281
Safety & Reliability

An AI That Checks Every Air Traffic Move Ahead of Time to Keep Planes Safe

Agent Mallard simulates candidate clearances through a realistic digital twin, finds coordinated, sector-wide conflict-free plans, and only issues commands that remain safe under modeled uncertainties.

Forward simulation combined with a lane-based airspace design turns complex lateral control into a d...

Credibility:
Must Read:
Jan 282
Agent Architectures

How an AI Can Work Forever on a Big Project Without Getting Lost

Externalizing the agent’s memory into files and keeping the reasoning context tiny makes AI agents far more stable over very long tasks — a 20‑billion‑parameter open model matched larger proprietary systems on a complex research benchmark.

Keeping long-term state in a structured workspace (files) and reconstructing a small, fixed reasonin...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

How Cheap Robots Learn to Gather Fast and Stay Together

Converting local bearing-only sensor readings into a small image and training decentralized policies with a global cohesion reward produces swarm behaviors that gather much faster while keeping groups connected better than classic analytic rules.

Encoding each agent’s local view as a 75×75 pixel image lets a convolutional network extract spatial...

Credibility:
Must Read:
Jan 28
Agent Architectures

Make Your Team of AI Helpers Faster by Letting One Controller Tune How They Talk

Centralize control of how AI agents talk and get resources so the platform can adapt messaging, batching, and routing in real time—raising throughput and meeting latency goals without changing each agent.

A controller that watches runtime metrics and can change communication granularity (from token-by-to...

Credibility:
Must Read:
Jan 28
Tool Use & Grounding

How Smart Programs Can Safely Spend Money on the Blockchain

Autonomous AI can act on blockchains usefully, but only if planning is separated from signing and every intent is checked by verifiable policy before funds move.

A clear progression of five integration patterns runs from read-only advisors to fully autonomous si...

Credibility:
Must Read:
Jan 28
Multi-Agent Systems

Find Strange Things on Camera Fast — and Understand Why

A two-agent cascade lets cheap, fast detectors handle routine camera frames and only sends hard cases to expensive language-capable vision models, cutting heavy checks threefold while keeping human-readable explanations.

A dual-agent design (one that reacts to alarms and one that continuously checks cameras) coordinates...

Credibility:
Must Read:
Jan 28
Benchmarks & Datasets

Live Test Shows AI Research Agents Still Fall Short on Financial Forecasts

AI research agents beat strong baselines on many forecasting tasks but still miss the mark on genuine forward-looking financial reasoning; a live, weekly benchmark exposes where they succeed and where they fail.

A live multi-agent evaluation system generates real, research-style forecasting tasks every week at ...

Credibility:
Must Read:
Jan 28