Steel

.@nibzard built a deep research agent on Steel. Then the evals taught him it was good at the wrong thing: beautiful overviews, weak exact answers. The fix was not another tool. It was routing, durability, and reading the failures. ↓

5 days. 5 drops. We're not saying what they are yet. Steel Launch Week starts Monday. @hussufo explains 👇

What's coming: ✦ identity, reliability & stealth ✦ new SDKs ✦ better pricing New drop every day, Mon–Fri. Follow along: steel.dev/launch-week

Full autopsy ↓ steel.dev/blog/claude-co…

Projects are live in Steel. ✦ Isolated namespaces inside one org ✦ Sessions, credentials, profiles, and API keys scoped per project ✦ One-way promotion from Development to Production Dev stops sharing a bucket with prod.

internally, we have been automating browsing agent recordings as a fun way of observing agents here you can enjoy a simple demo of claude code actions overlayed in sync with a browser session

Changelog #29: docs.steel.dev/changelog/chan… or, come hang in our discord: discord.gg/steel-dev

What's new @ Steel - Changelog #029 ✦ Projects + project-scoped API keys and sessions ✦ Sessions can now trust custom CAs with caCertificates ✦ v1/scrape markdown overhaul: better extraction, richer metadata, full-page fallback ✦ Plus live viewer upgrades, browser pooling & more Link below ↓

Nothing to migrate. Every org now has a Default project with your resources and keys already in it. Your code keeps working. Full write-up. steel.dev/blog/introduci…

We held off on this longer than most platforms. Our honest opinion was simplicity trumps complexity: one org, one bucket, nothing to configure. Then one agent became a fleet, and simple scripts became workflows carrying their own credentials and profiles. One bucket doesn't scale with agents.

Not sure if I wrote this, or if this one wrote me.

Steel mattered because the research needed browsers, not fetches. The useful web renders late, redirects, blocks scrapers, or hides content behind behavior. Plain fetch sees a thinner world. Deep research agents improve when every failed run leaves enough evidence to inspect, resume, and turn into the next change. steel.dev/blog/durable-r…

Pi from @badlogicgames for the agent loop, Absurd by @mitsuhiko for durable execution, Steel for browser sessions, Postgres for checkpoints. The real unlock was making failure resumable. Every model message checkpoints. Notes, URLs, claims, and source ledgers rebuild from the transcript. Crashes became annoying instead of existential.

.@nibzard built a deep research agent on Steel. Then the evals taught him it was good at the wrong thing: beautiful overviews, weak exact answers. The fix was not another tool. It was routing, durability, and reading the failures. ↓

The "deep research" workflow bundled in Claude Code isn't deep. It's wide. Niko pulled Claude Code's dynamic workflow to find exactly that shape. ↓

@nibzard depth is not free

claude code /deep-research isn't deep. it's wide. your question gets sprayed across 5 parallel searches, the pile gets summarized, done. no agent ever reads a result and forms a sharper question from it. i pulled the workflow out of its binary to inspect it. ↓

@Teknium @Teknium check our boy @nibzard’s PR

@steeldotdev I think you forgot hermes good sir

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.