Latest posts
A demo that looked like half an hour of work — record yourself publishing an article, replay it with a different one — broke twenty-odd times across Zhihu and Xiaohongshu. This post walks every trap in the order we hit it: noopener new tabs, frozen renderers, programmatic-click poisoning, SPA route echoes, closed shadow-DOM dead zones, click actionability, the network-idle race — up to the server-side validation that only evidence could pin down. With a diagnostic decision tree, and one honest lesson about chasing ghosts.
- Why I Built Another Open-Source Agent
The origin story of Octo, a self-hosted AI agent in a single Go binary — why I built it, the design decisions behind it, and what it deliberately won't do.
- Octo Onboarding Series (10): Record & Replay in Practice — Record Your Actions Once, Let octo Replay Them
Record a browser workflow you'd otherwise do by hand, distill it into a replayable, self-healing script — the next time it comes up, octo just clicks through it.
- Octo Onboarding Series (9): Browser in Practice — Hand octo Your Own Browser
The browser tool isn't another headless framework — it attaches to your real, already-logged-in Chrome, and a click or form submit is treated as a real action that needs approval.
- Octo Onboarding Series (8): Goal in Practice — Set a Standing Objective and Let It Find Idle Time to Push On
Migrating an API from v1 to v2 across every call site isn't a one-message task — it takes several passes. That's what goal is for.
- Octo Onboarding Series (7): Workflow in Practice — Get Several Agents Working in Parallel
A multi-angle code review doesn't have to run one pass at a time — a workflow script reviews from three angles at once, then merges them into one verdict.
- Octo Onboarding Series (6): Putting It All Together — A Real, Running Weekly Report
Cron fires it, MCP fetches the data, a skill builds the Excel file, IM delivers it — the previous five posts wired into one automation that actually runs every week.
- Octo Onboarding Series (5): Cron in Practice — Scheduled Tasks That Run Whether You're There or Not
Every morning at 9, without you asking, octo has already summarized what happened in the repo yesterday and sent it to you.
- Octo Onboarding Series (4): Loop in Practice — Get octo to Watch Something for You
You pushed a PR, CI takes a few minutes, and you don't want to refresh the page every two minutes — let it watch and tell you when it's green.
- Octo Onboarding Series (3): MCP in Practice — Connect GitHub and Let octo Triage Your Issues
Wiring up an external tool takes one JSON file — once it's connected, octo can go read your GitHub issues for you.
- Octo Onboarding Series (2): Skills in Practice — Generate an Excel Report in One Sentence
No openpyxl to learn, no functions to remember — describe what you want, and the built-in office-xlsx skill takes it from there.
- Octo Onboarding Series (1): Install It, Say Your First Word to It
Terminal or browser — pick one path, finish it, and get octo to do something real for you in five minutes.
- octo-agent Deep Dive: The Genuinely Hard Parts of an AI Agent System
Starting from an agent loop of only a few hundred lines, this post dissects the ten mechanisms that keep octo-agent stable, fast, and cheap: the built-in tool contract, context compaction, prompt caching, three lifetimes of memory, the robustness design of octo serve, workflow orchestration, a self-healing browser, the permission system, the trash can, and batteries included. Plus the one line running through them all: whatever a mechanism can guarantee, never leave to the model's self-discipline.