Published: 2026-09-01
A fully local agent with tools, in about 40 lines: Ollama plus Pydantic AI
Running Nex-N2.5 Mini on two H100s: the SGLang container setup, and an honest benchmark read
Semantic grep cut agent tool calls 58% and input tokens 47% in the project's own benchmarks
The instruction ceiling moved 10x in a year: 200 rules became 2,000
How LinkedIn made coding agents work on 1,000+ internal repos without fine-tuning
Two context approaches that look right and stall: the curated-context trap and the MCP plateau
OpenClaw 2.0 from a clean install: the Lab toggles, dreaming memory and Telegram pairing
Chapters / key moments (click to jump — plays here on the page)
The install went from roughly fifteen terminal questions to two or three, with the rest moved into a conversation with the agent — and on first run it scans your machine and wires up what it finds. In this run it detected a locally served model and an existing Claude subscription without either being configured by hand. The section worth opening first is Lab, which holds the pre-release toggles including tool-loop detection, code mode, swarm and tool search. One thing to know before you upgrade rather than install fresh: the project's own update guidance acknowledges breakage and suggests using a coding agent to work through it.
Source video
"OpenClaw 2.0 Is Finally Here — But Is It Worth Using?" by Bart Slodyczka — Watch on YouTube →
Gotchas & Caveats
- Upgrading from 1.x is the risky path, not the install. Many users reported breakage, and the official update guidance is to use another coding harness to work through failures. This run was a deliberate clean install for that reason.
- The reviewer had used it for one day at the time of recording and says so repeatedly — the memory dreaming, skills workshop and automations had not yet accumulated anything to show. Treat the feature descriptions as verified and the quality judgements as pending.
- Some settings toggles did not behave obviously — enabling a channel did not surface a setup window, and the setup turned out to live on a different settings page.
- The in-UI browser opened but did not respond to input in this run.
Key Takeaways
- The install is front-loaded no longer. Two or three questions, then you finish setup by talking to the agent — connecting Telegram, tools and plugins from a conversation rather than a terminal prompt.
- First-run scanning works. With approval, it looks for anything on your machine it can plug into; here it found a locally served model and an existing Claude subscription and connected both. Note that it drives the Claude CLI directly rather than asking you to sign in.
- Memory now runs in three stages rather than pinning everything you say. A light stage sifts recent conversations for candidates, a REM stage groups them into themes, and a deep stage is what actually writes to the memory file. Dreaming is on by default and appears in the automations list.
- The Lab section holds the pre-release toggles: tool-loop detection (rolling history guards that warn or block repeated tool calls when an agent stops progressing), code mode, swarm, and tool search. Tool-loop detection is the one to enable if you run smaller local models, which loop more.
- Automations ships with three defaults: a heartbeat every 30 minutes, the memory dreaming process, and a skill collection review.
- The skills workshop mines your own history. It reviews substantial sessions newest-first looking for repeated workflows and recovery patterns worth turning into a skill — the same idea Hermes made its name on.
- Telegram pairing is a complete flow now: paste a bot token, choose the DM policy (pairing), start the bot, copy the pairing code it returns, then approve the pending request from the channels settings page. Working in a couple of minutes.
- A browser is built into the UI, and there are native app extensions including a Chrome side-panel extension. The Mac and browser apps share the same UI; the browser app felt marginally smoother in this run.
More OpenClaw & Claude Code news
A fully local agent with tools, in about 40 lines: Ollama plus Pydantic AI
Running Nex-N2.5 Mini on two H100s: the SGLang container setup, and an honest benchmark read
Semantic grep cut agent tool calls 58% and input tokens 47% in the project's own benchmarks
The instruction ceiling moved 10x in a year: 200 rules became 2,000
How LinkedIn made coding agents work on 1,000+ internal repos without fine-tuning
Two context approaches that look right and stall: the curated-context trap and the MCP plateau