Steel

Humans use Chrome, Agents use Steel.

Feed this to your OpenClaw or Hermes:
curl -fsS | sh

[Steel @steeldotdev]

The team's been deep on the different shapes a browsing agent can take.

Agents fail in three ways on long runs: they quit early, grade their own work on a curve, and forget the rules after a compaction.

We'd been fighting them in a browser for months.

Three things we'd add from here 👇

[Niko(la) @nibzard]

https://t.co/dGW2a03ica

[Steel @steeldotdev]

Today we are launching Steel Skills.

Five agent skills for the web. Install one or the whole set. Runs in Claude Code, Cursor, Codex, opencode, Pi, or any compatible agent.

[Steel @steeldotdev]

Install one skill, run one real flow, tell us one blocker and one pass.

Start here → https://t.co/8fcBEeiipL

[Steel @steeldotdev]

Today we are launching Steel Skills.

Five agent skills for the web. Install one or the whole set. Runs in Claude Code, Cursor, Codex, opencode, Pi, or any compatible agent.

[Steel @steeldotdev]

List the catalog with @vercel skills or use Steel CLI skills command.

`npx skills add steel-dev/skills --list`

[Steel @steeldotdev]

A safety net for the runs where your code or agent never gets the chance to clean up.

Read more about it in this writeup by @junhssss

https://t.co/Pc0S2YLHzX

[Steel @steeldotdev]

Stop paying for a browser after the work is done or your agent dies.

Steel sessions now take an inactivity timeout: if the agent stops driving the browser for the window you set, Steel releases it and the meter stops.

Read more ↓

[Steel @steeldotdev]

The honest fix was always "release the session yourself." And you should.

Now you can just set the timeout, and Steel handles the cleanup for you.

Just set the window longer than the longest gap you expect between commands.

[Steel @steeldotdev]

Most of what you ship to an agent gets compressed before it acts — docs, SDKs, blog posts, distilled by the model first.

Errors are the exception. They reach the agent intact.

We've been rebuilding ours around that. ↓

[Steel @steeldotdev]

The team's been deep on the different shapes a browsing agent can take.

Agents fail in three ways on long runs: they quit early, grade their own work on a curve, and forget the rules after a compaction.

We'd been fighting them in a browser for months.

Three things we'd add from here 👇

[Steel @steeldotdev]

@0xbosta How we think about talking to an agent now — three channels:
✦ hard error: it failed, here's the fix
✦ soft warning: it worked, but you'll regret it
✦ agent notes: ambient per-site context

Full writeup by @0xbosta ↓
https://t.co/DilGBvMrEt

[Steel @steeldotdev]

Most of what you ship to an agent gets compressed before it acts — docs, SDKs, blog posts, distilled by the model first.

Errors are the exception. They reach the agent intact.

We've been rebuilding ours around that. ↓

[Steel @steeldotdev]

Changelog #027: https://t.co/lBBtQzCSBm

or, come hang in our discord: https://discord. gg/steel-dev

[Steel @steeldotdev]

What's new @ Steel - Changelog #027

✦ New Agent Traces docs: overview, timeline + exports, and the API
✦ Leaderboard refresh — filterable /results index and cleaner benchmark pages
✦ More benchmark entries, plus tooling to find new results worth listing
✦ Plus docs-delivery fixes and browser-seconds metering under the hood

Link below ↓

[Steel @steeldotdev]

Latest: @AnthropicAI Claude Opus 4.8 is sitting #1 on OSWorld @XLangNLP at 83.4%, above the human baseline. Added the day it shipped.

https://t.co/41hgmZzPuW

[Steel @steeldotdev]

Claude Opus 4.8 now ranks #1 on the OSWorld benchmark with an 83.4% score.

We just added it to the Steel leaderboard.

Congrats to the Anthropic team.

[Steel @steeldotdev]

Browser-agent benchmarks are getting crowded, stale, and hard to compare across.

We are collecting the benchmarks that actually matter for browser and computer-use agents, so you don't have to chase them down.

(Just rebuilt the whole leaderboard. Live now.)

[Steel @steeldotdev]

See where every browser agent actually ranks, and how each number was scored.

https://t.co/uw1i6kczZT

Would love to see benchmark maintainers sanity-check. See a result we're missing, or running an agent with a real score? Add it.

@shuyanzh36 @SWEbench @jyangballin @_carlosejimenez @XLangNLP @TianbaoX @taoyds @OpenAI @_jasonwei