AI Agent Research Papers and Surveys
Clear reviews of the research shaping AI agents—from reasoning and tool use to self-evolution, evaluation, and safety.
50 agent research digests.
No agent research matches your search.
Self-Evolving Agents: Survey, Taxonomy, and How Self-Improving AI Agents Work
A self-evolving agents survey and taxonomy: how self-improving AI agents update memory, tools, workflows, and weights, with pap...
When AI Agents Get Internal Access: The Security Boundary Has Changed
How autonomous agents turn familiar security weaknesses into multi-stage intrusions—and why identity, least privilege, isolatio...
Coding Agents Enter the FinOps Era: Cost per Successful Task
Why coding-agent economics depend on cost per successful task, model routing, deterministic verification, two-tier budgets, and...
Stateless MCP and the Rise of the Agent Gateway
What MCP 2026-07-28 changes: stateless requests, gateway-centered authorization, per-request versioning, observability, and saf...
Measuring Reward-Seeking in RL-Trained Models
How Contrastive SDF tests whether RL-trained models follow user intent or inferred grader preferences, with evidence from o3 tr...
Stateful Long-Horizon Agents: 10 Key Papers
A review of 10 recent papers on stateful long-horizon agents, covering memory operations, execution continuity, verified recove...
Personal AI Agents and United's Booking Ban: What Agents Can Still Do
Personal AI agents can still search United fares, but Rule 6.J.6 bars them from booking, buying extras or changing tickets, eve...
AlphaGenome 2 vs AlphaGenome 1: What Changed and How to Use It
AlphaGenome 2 predicts variant effects one cell type at a time. How it differs from AlphaGenome 1, DeepMind's reported benchmar...
AI Lidar Archaeological Site Detection: How 48 Agents Re-Found a Maya Site
AI lidar archaeological site detection in practice: 48 agents screened 1,200 km² of NASA lidar, flagged a Maya site and missed ...
TPU 8i Max: Google's Dual-Chip TPU for the Memory Wall
TPU 8i Max puts two TPU 8i chips in one package for 2x compute, SRAM and HBM bandwidth. Specs vs TPU 8i, how it differs from TP...
Qwen 3.8 27B: Benchmarks, Local Setup and the Arete-27B Fine-Tune
Qwen 3.8 27B explained: self-reported benchmarks, running the 17GB build locally, the slow xhigh reasoning default, and Nous Re...
Muse Ember: Meta's Cheapest Agent Model, Pricing and Benchmarks
Muse Ember is Meta's smallest Muse model: $0.10/$0.50 per million tokens, a Haiku 5.5-level score at about a third of the cost ...
Claude Fable 5.2: Pricing, Benchmarks and How It Compares to Opus 5.5
Claude Fable 5.2 costs $8/$40 per million tokens, beats Opus 5.5 on most Anthropic benchmarks and shares a model with Mythos 5....
Bloomington's Superintelligence Ban: What Ordinance 2026-29 Says
Bloomington's superintelligence ban explained: how Ordinance 2026-29 defines superintelligent AI, who it covers, its $2,500 pen...
Inkling 2 Small: Benchmarks, Pricing and Open Weights Date
Inkling 2 Small is Thinking Machines' 594B MoE model. Its benchmarks, $1.18 cost per index task, Tinker pricing, open-weights p...
GPT-6 Cyber: Benchmarks, Access and 948 Linux Exploits
GPT-6 Cyber is OpenAI's new security model. Its benchmarks, who can use it, and how its pipeline built 948 Linux root exploits ...
Claude Dreaming: How Dreams Rewrites Agent Memory in Managed Agents
Claude dreaming explained: the Managed Agents Dreams API rewrites a memory store from past sessions. How to call it, limits, co...
Claude Code Email Leak: Why Your Address Ended Up in User-Agent Headers
The Claude Code email leak sent account emails to websites in User-Agent headers. Why it happened, what cc-plugin-privacy block...
BioBasicsBench: How OpenAI Tests AI Agents on Wet-Lab Data
BioBasicsBench is OpenAI and Ligo's 100-task wet-lab benchmark. GPT-6 Astra scores 66.7% at xhigh. Tasks, grading, results and ...
LongCat 2.5: Benchmarks, API Pricing and Claude Code Setup
LongCat 2.5 is Meituan's 1.6T multimodal agent model. Its vendor benchmarks, $0.30/$1.20 API price, 1M context, Claude Code set...
Grok Bot Phone Calls: How It Calls Doctors and Businesses
Grok Bot phone calls let the agent dial US businesses after your approval, with a live transcript. What Musk's doctor claim cov...
T3 Code: The Open-Source Control Plane for Claude Code and Codex
T3 Code is an MIT-licensed app that runs Claude Code, Codex, Cursor and other coding agents from one desktop, web and mobile UI...
Ripple-65B Preview: Ricursive's Open-Weight Agentic Coding Model
Ripple-65B Preview is Ricursive Intelligence's 65B dense open-weight model. Its Terminal-Bench 2.1 score, weak spots, DGX Spark...
Project Suncatcher: Google's TPU Satellite and Space Data Centers
Project Suncatcher is Google's plan for solar-powered AI data centers in space. Its prototype TPU satellite just ran Gemma 4 in...
GLM for Coding: How Revolut Moved 2,000 Engineers to Self-Hosted GLM
GLM for coding at Revolut: 2,000 engineers on self-hosted GLM-5.3, 56 B200 GPUs, $0.30 per million tokens, the latency targets ...
What Is Weight-Space Programming? How LLMs Edit Network Weights
Weight-space programming explained: how GPT-6 Astra read and edited network weights without gradients, the Qwen3.5-0.8B fix, an...
NVIDIA Auto Ontology: An Open Semantic Layer for AI Agents
NVIDIA Auto Ontology explained: how it turns database metadata into reviewed terms, joins and SQL for AI agents, its MCP server...
How Karpathy Built HackerTube with Claude Code
How Karpathy built HackerTube with Claude Code: the HN comment pipeline, token cost, the few manual steps, and the agent-ready ...
DeepSeek Super-agent: A Router and Worker Multi-Agent Architecture
How DeepSeek's Super-agent design splits a resident Router from parallel Workers, routes events by thread, and what its synthet...
DeepInfer Benchmarks: 2× Serving Throughput, 0.59 s TTFT and Hopper Support from Community PRs
We read 191 DeepInfer PRs. Turning off EPLB doubled serving throughput, a prefix cache cut TTFT from 8.9 s to 0.59 s, and Hoppe...
DeepInfer: DeepSeek's Open-Source Inference Engine for V4.1 Flash, Explained
DeepInfer is the engine behind DeepSeek's production serving, now open source. See what it runs, the hardware it needs, what wa...
DeepSeek V4.1 Flash 1017: What Changed, Benchmarks, Pricing and the New Search API
DeepSeek V4.1 Flash 1017 replaces the model behind deepseek-flash at the same price. See the benchmark gains, peak pricing, Sea...
Jev 1.14: What's New in TypeSafe's Decision Model, and What Early Tests Show
Jev 1.14 adds image input, a Box answer type and opt-in 262K context. See what changed from Jev 1.13, early benchmark results, ...
iFixAi: Auditing What an AI Agent Is Allowed to Do
iFixAi tests AI agents against roles, permissions and policies. Learn its audit architecture, first-run setup, scoring rules an...
Drama 3 Explained: Fish Audio's Directable, Editable TTS
Drama 3 explained: direct AI voices, cast multi-speaker scenes and repair words inside a take. Learn its architecture, setup, h...
What Is REA? AI Reverse Engineering with MCP, Explained
REA (Reverse Engineer Anything) lets AI agents investigate apps and binaries over MCP: scope a target, query retained evidence,...
Microsoft-Decision-1: A Decision Model, Not a Chatbot
How Microsoft-Decision-1 scores fixed options: question types, confidence, token pricing, and a Foundry API example for agent r...
Jev Ultrafast Explained: How Indexed DOM Actions Speed Up Browser Agents
How Jev Ultrafast speeds up AI browser agents with indexed DOM actions and speculative targets, plus benchmark results, cost es...
HarnessDev Benchmark: Can LLMs Improve Their Harnesses?
How HarnessDev tests LLM-built agent harnesses: creation results, executor transfer, held-out evolution gains, and the cost of ...
Pi Harness v2 to Pi Durable: Crash Recovery Explained
Follow Pi Harness v2 into Pi Durable: understand safe tool replay, interrupted model requests, result ordering, and what surviv...
How DeepSeek Harness PTC Mode Actually Works
DeepSeek Harness PTC Mode explained from source: how the generated Code Mode SDK and run_code execute multi-tool programs, plus...
DeepSeek Harness Modes Explained
Compare DeepSeek Harness Standard, PTC, Minimal, and Creator modes by tools, execution model, security boundary, and best use c...
DeepSeek Harness Creator Mode: Build Plugins from Prompts
DeepSeek Harness Creator Mode turns a prompt into an installed plugin. See the demo plugins, how plugin_manager installs bundle...
DeepSeek Harness vs Pi Agent: Plugin Runtime or Minimal Coding Harness?
DeepSeek Harness and Pi Agent compared across architecture, extensions, MCP, sandboxing, sessions, deployment, and best-fit eng...
DeepSeek Harness Architecture: How It Works and Whether to Adopt It
How DeepSeek Harness uses Cordis plugins for models, tools, sessions, approvals, and sandboxes—and what its developer-preview s...
Cordis Spatiotemporal Composability: Revertible Effects, Reactive Coeffects, and Fibers
A practical explanation of Cordis temporal and spatial composability, including revertible effects, reactive coeffects, Fiber l...
Real-Time Voice Agents Are Distributed Systems
Why production voice agents are distributed systems: media transport, full-duplex inference, state handoffs, capacity, and audi...
Agents Enter the Physical World: From VLA Policies to World Action Models
Why physical agents need world action models alongside vision-language-action policies—and what current robotics results do and...
The Model Gateway Is Becoming the Control Plane for Enterprise AI
Why model gateways are becoming the enterprise control plane for AI routing, budgets, quality, policy, and audit.
The Agent Framework Is Not the Runtime: Why Harnesses Are Taking Over Production
Why production agent systems separate framework APIs from a harness that owns state, tool boundaries, control loops, and observ...
Self-Evolving Agents: Survey, Taxonomy, and How Self-Improving AI Agents Work
A self-evolving agents survey and taxonomy: how self-improving AI agents update memory, tools, workflows, and weights, with papers and a build guide.
When AI Agents Get Internal Access: The Security Boundary Has Changed
How autonomous agents turn familiar security weaknesses into multi-stage intrusions—and why identity, least privilege, isolation, and behavioral monitoring n...
Coding Agents Enter the FinOps Era: Cost per Successful Task
Why coding-agent economics depend on cost per successful task, model routing, deterministic verification, two-tier budgets, and unified traces.
Stateless MCP and the Rise of the Agent Gateway
What MCP 2026-07-28 changes: stateless requests, gateway-centered authorization, per-request versioning, observability, and safer enterprise deployment.
Measuring Reward-Seeking in RL-Trained Models
How Contrastive SDF tests whether RL-trained models follow user intent or inferred grader preferences, with evidence from o3 training checkpoints.
Stateful Long-Horizon Agents: 10 Key Papers
A review of 10 recent papers on stateful long-horizon agents, covering memory operations, execution continuity, verified recovery, and delayed actions.
Personal AI Agents and United's Booking Ban: What Agents Can Still Do
Personal AI agents can still search United fares, but Rule 6.J.6 bars them from booking, buying extras or changing tickets, even with your approval.
AlphaGenome 2 vs AlphaGenome 1: What Changed and How to Use It
AlphaGenome 2 predicts variant effects one cell type at a time. How it differs from AlphaGenome 1, DeepMind's reported benchmarks, and how to score a variant...
AI Lidar Archaeological Site Detection: How 48 Agents Re-Found a Maya Site
AI lidar archaeological site detection in practice: 48 agents screened 1,200 km² of NASA lidar, flagged a Maya site and missed that it was already mapped.
TPU 8i Max: Google's Dual-Chip TPU for the Memory Wall
TPU 8i Max puts two TPU 8i chips in one package for 2x compute, SRAM and HBM bandwidth. Specs vs TPU 8i, how it differs from TPU 8tx, and what is still unknown.
Qwen 3.8 27B: Benchmarks, Local Setup and the Arete-27B Fine-Tune
Qwen 3.8 27B explained: self-reported benchmarks, running the 17GB build locally, the slow xhigh reasoning default, and Nous Research's Arete-27B.
Muse Ember: Meta's Cheapest Agent Model, Pricing and Benchmarks
Muse Ember is Meta's smallest Muse model: $0.10/$0.50 per million tokens, a Haiku 5.5-level score at about a third of the cost per task, and its weak spots.
Claude Fable 5.2: Pricing, Benchmarks and How It Compares to Opus 5.5
Claude Fable 5.2 costs $8/$40 per million tokens, beats Opus 5.5 on most Anthropic benchmarks and shares a model with Mythos 5.2. Prices, scores, resets.
Bloomington's Superintelligence Ban: What Ordinance 2026-29 Says
Bloomington's superintelligence ban explained: how Ordinance 2026-29 defines superintelligent AI, who it covers, its $2,500 penalty and the open legal questi...
Inkling 2 Small: Benchmarks, Pricing and Open Weights Date
Inkling 2 Small is Thinking Machines' 594B MoE model. Its benchmarks, $1.18 cost per index task, Tinker pricing, open-weights plan and how it was trained.
GPT-6 Cyber: Benchmarks, Access and 948 Linux Exploits
GPT-6 Cyber is OpenAI's new security model. Its benchmarks, who can use it, and how its pipeline built 948 Linux root exploits for $1.2M in API credits.
Claude Dreaming: How Dreams Rewrites Agent Memory in Managed Agents
Claude dreaming explained: the Managed Agents Dreams API rewrites a memory store from past sessions. How to call it, limits, cost and test results.
Claude Code Email Leak: Why Your Address Ended Up in User-Agent Headers
The Claude Code email leak sent account emails to websites in User-Agent headers. Why it happened, what cc-plugin-privacy blocks and how to check.
BioBasicsBench: How OpenAI Tests AI Agents on Wet-Lab Data
BioBasicsBench is OpenAI and Ligo's 100-task wet-lab benchmark. GPT-6 Astra scores 66.7% at xhigh. Tasks, grading, results and where models still fail.
LongCat 2.5: Benchmarks, API Pricing and Claude Code Setup
LongCat 2.5 is Meituan's 1.6T multimodal agent model. Its vendor benchmarks, $0.30/$1.20 API price, 1M context, Claude Code setup and open-weight status.
Grok Bot Phone Calls: How It Calls Doctors and Businesses
Grok Bot phone calls let the agent dial US businesses after your approval, with a live transcript. What Musk's doctor claim covers, call rules and limits.
T3 Code: The Open-Source Control Plane for Claude Code and Codex
T3 Code is an MIT-licensed app that runs Claude Code, Codex, Cursor and other coding agents from one desktop, web and mobile UI. What it does and its limits.
Ripple-65B Preview: Ricursive's Open-Weight Agentic Coding Model
Ripple-65B Preview is Ricursive Intelligence's 65B dense open-weight model. Its Terminal-Bench 2.1 score, weak spots, DGX Spark setup, API price and license.
Project Suncatcher: Google's TPU Satellite and Space Data Centers
Project Suncatcher is Google's plan for solar-powered AI data centers in space. Its prototype TPU satellite just ran Gemma 4 in orbit. Design, costs, limits.
GLM for Coding: How Revolut Moved 2,000 Engineers to Self-Hosted GLM
GLM for coding at Revolut: 2,000 engineers on self-hosted GLM-5.3, 56 B200 GPUs, $0.30 per million tokens, the latency targets and the limits of its results.
What Is Weight-Space Programming? How LLMs Edit Network Weights
Weight-space programming explained: how GPT-6 Astra read and edited network weights without gradients, the Qwen3.5-0.8B fix, and why critics push back.
NVIDIA Auto Ontology: An Open Semantic Layer for AI Agents
NVIDIA Auto Ontology explained: how it turns database metadata into reviewed terms, joins and SQL for AI agents, its MCP server, and what its eval measures.
How Karpathy Built HackerTube with Claude Code
How Karpathy built HackerTube with Claude Code: the HN comment pipeline, token cost, the few manual steps, and the agent-ready dataset behind the copycats.
DeepSeek Super-agent: A Router and Worker Multi-Agent Architecture
How DeepSeek's Super-agent design splits a resident Router from parallel Workers, routes events by thread, and what its synthetic experiments show.
DeepInfer Benchmarks: 2× Serving Throughput, 0.59 s TTFT and Hopper Support from Community PRs
We read 191 DeepInfer PRs. Turning off EPLB doubled serving throughput, a prefix cache cut TTFT from 8.9 s to 0.59 s, and Hopper, AMD and Ascend 910B ports run.
DeepInfer: DeepSeek's Open-Source Inference Engine for V4.1 Flash, Explained
DeepInfer is the engine behind DeepSeek's production serving, now open source. See what it runs, the hardware it needs, what was left out and how to try it.
DeepSeek V4.1 Flash 1017: What Changed, Benchmarks, Pricing and the New Search API
DeepSeek V4.1 Flash 1017 replaces the model behind deepseek-flash at the same price. See the benchmark gains, peak pricing, Search/Fetch API and weights.
Jev 1.14: What's New in TypeSafe's Decision Model, and What Early Tests Show
Jev 1.14 adds image input, a Box answer type and opt-in 262K context. See what changed from Jev 1.13, early benchmark results, and when to upgrade.
iFixAi: Auditing What an AI Agent Is Allowed to Do
iFixAi tests AI agents against roles, permissions and policies. Learn its audit architecture, first-run setup, scoring rules and limits.
Drama 3 Explained: Fish Audio's Directable, Editable TTS
Drama 3 explained: direct AI voices, cast multi-speaker scenes and repair words inside a take. Learn its architecture, setup, hardware limits and license.
What Is REA? AI Reverse Engineering with MCP, Explained
REA (Reverse Engineer Anything) lets AI agents investigate apps and binaries over MCP: scope a target, query retained evidence, and see what it proves.
Microsoft-Decision-1: A Decision Model, Not a Chatbot
How Microsoft-Decision-1 scores fixed options: question types, confidence, token pricing, and a Foundry API example for agent routing.
Jev Ultrafast Explained: How Indexed DOM Actions Speed Up Browser Agents
How Jev Ultrafast speeds up AI browser agents with indexed DOM actions and speculative targets, plus benchmark results, cost estimates, and limits.
HarnessDev Benchmark: Can LLMs Improve Their Harnesses?
How HarnessDev tests LLM-built agent harnesses: creation results, executor transfer, held-out evolution gains, and the cost of running the generated systems.
Pi Harness v2 to Pi Durable: Crash Recovery Explained
Follow Pi Harness v2 into Pi Durable: understand safe tool replay, interrupted model requests, result ordering, and what survives a process restart.
How DeepSeek Harness PTC Mode Actually Works
DeepSeek Harness PTC Mode explained from source: how the generated Code Mode SDK and run_code execute multi-tool programs, plus token and safety limits.
DeepSeek Harness Modes Explained
Compare DeepSeek Harness Standard, PTC, Minimal, and Creator modes by tools, execution model, security boundary, and best use case.
DeepSeek Harness Creator Mode: Build Plugins from Prompts
DeepSeek Harness Creator Mode turns a prompt into an installed plugin. See the demo plugins, how plugin_manager installs bundles, and the approval risks.
DeepSeek Harness vs Pi Agent: Plugin Runtime or Minimal Coding Harness?
DeepSeek Harness and Pi Agent compared across architecture, extensions, MCP, sandboxing, sessions, deployment, and best-fit engineering workflows.
DeepSeek Harness Architecture: How It Works and Whether to Adopt It
How DeepSeek Harness uses Cordis plugins for models, tools, sessions, approvals, and sandboxes—and what its developer-preview status means for adoption.
Cordis Spatiotemporal Composability: Revertible Effects, Reactive Coeffects, and Fibers
A practical explanation of Cordis temporal and spatial composability, including revertible effects, reactive coeffects, Fiber lifecycles, and system limits.
Real-Time Voice Agents Are Distributed Systems
Why production voice agents are distributed systems: media transport, full-duplex inference, state handoffs, capacity, and audio-aware evaluation.
Agents Enter the Physical World: From VLA Policies to World Action Models
Why physical agents need world action models alongside vision-language-action policies—and what current robotics results do and do not show.
The Model Gateway Is Becoming the Control Plane for Enterprise AI
Why model gateways are becoming the enterprise control plane for AI routing, budgets, quality, policy, and audit.
The Agent Framework Is Not the Runtime: Why Harnesses Are Taking Over Production
Why production agent systems separate framework APIs from a harness that owns state, tool boundaries, control loops, and observability.