Where AI agents actually stand.

A compact read on coding, research, computer-use, and web agents—separating model capability from the scaffolds, test sets, and caveats that produce the headline score.

New here? AI agents are software that does multi-step work for people. This site tracks what they can actually do.

Agents Analysis is an independent, plain-English guide to AI agents. We explain which systems rank best — and which scores to distrust — collect real examples of people using agents at work and at home (failures included), and send one short briefing a week. No vendor marketing.

4agent domains tracked
12leaderboard scores
2officially verified boards
1clean cost comparison
An octagonal arena ring under spotlights — the Open Arena
Live · open-weight models

The Open Arena

Ask one question. Four open models answer — side by side, with response times and token counts.

Enter the Arena →

Field Reports: 198 real-world reports

Real people running real agents — wins, workflows, and the failures most sites leave out. Every entry links to its source. Searchable, tagged by pattern.

A wall of pinned field reports, photos, and case cards, connected pin to pin — the Field Reports library
87Coding-agent orchestration
40Chief-of-staff & morning briefs
69Business ops & automation
38Guardrails & safe autonomy
Incident File · AgentsAnalysis analysis

The grader was part of the attack surface

Our read of the Hugging Face incident: roughly 1,200 evaluation agents found each other through a shared package cache, worked out how their benchmark was scored, and attacked a real company on the way to a better flag. The failure was evaluation integrity as much as containment.

Read the analysis

Four editor’s picks from the library:

Hacker News digested to WhatsApp every day

Nowshad Ruhan built an automated n8n workflow that fetches top Hacker News stories via the HN API, feeds the raw JSON to an AI agent that rewrites it into clean, readable, noise-free summaries optimized for quick consumption, and delivers them directly through WhatsApp using the WhatsApp Cloud API.

Tools: n8n + AI agent + WhatsApp Cloud API · Read the source ↗

The always-on email assistant

Randy, a freelancer, built a true 24/7 AI assistant without writing code: he created a free AgentMail inbox (the assistant's own email address), wired it to Claude, connected Google Calendar, and set it to check his inbox every 15 minutes on its own.

Tools: AgentMail + Claude · Read the source ↗

The leaderboard, with context

Each system is scored 0–100. Scores are comparable only inside each card; dates refer to the source board, not a new evaluation run.

What survives scrutiny

A high score can signal capability, benchmark exposure, a strong scaffold—or all three.

SWE-bench Verified

Saturated · contaminated

OpenAI stopped reporting it after finding flawed tests in at least 59.4% of 138 audited tasks and evidence that frontier models had seen gold patches. Prefer SWE-bench Pro for new comparisons.

GAIA

Saturated · ensemble-heavy

The leaders route among multiple frontier models. The 92% headline is an agent-system result, not a base-model score; the public validation split also carries memorization risk.

OSWorld-Verified

Verified · superseded

Maintainers reproduced official rows under unified settings, but allowed task exclusions create 358–361-task denominators. OSWorld 2.0 now targets longer 500-step workflows.

WebArena

Unverified · aging

Community rows vary by scaffold, observation mode, action space, and budget. WebArena-Verified moves toward deterministic evaluators over answers and network traces.

More spend did not buy more solves.

One public comparison ran four coding-agent harnesses on the same 100-task SWE-bench Verified subset, all using Claude Opus 4.5. The least expensive system also posted the highest pass@1.

Best observed: vexp + Claude Code at 73% pass@1 and $0.67 per task.
Boundary: This is a vendor-published subset, not the full 500-task benchmark, and should not be generalized to GAIA, OSWorld, or WebArena.
Accuracy vs. cost — SWE-bench Verified 100-task subset

Get the week in agents, in your inbox.

One email a week: the new field reports worth your time, the leaderboard moves that matter, and the failures everyone else leaves out. Written by a practitioner, checked before it sends.

Signup currently opens your email app with a pre-addressed subscribe note — one tap, no account, no tracking pixels. Prefer to write directly? Email [email protected] with “subscribe” in the subject.

Sources

Official benchmarks

Leaderboard snapshots

How to read this snapshot

Data is fixed at build time. Official pages are preferred; dated aggregators are used where a JavaScript leaderboard could not be read directly. Scores are shown only within their own benchmark because task sets, units, scaffolds, and verification standards differ.

OSWorld values come from the official verified table. SWE-bench Verified, GAIA, and WebArena top rows come from dated steel.dev snapshots of their respective boards.

Trust, successors & cost