It does more than you think
In one session the agent reads and writes files, runs commands, browses the web, delegates to subagents, orchestrates workflows, loads skills, drives long-running goals — and even mounts new capabilities onto itself.
3
entry modes
2
doc languages
100%
plugin-based
Watch one real agent session
Read & write files
Read, create and edit workspace files, strictly inside the workspace you select.
Translate the README into Chinese and save it as README.zh.md
Run commands
Run shell commands under your approval: tests, dependencies, builds.
Run pnpm test and fix every failing case
Delegate to subagents
Fan independent tasks out to subagents in parallel, then aggregate.
Let three subagents review three modules and collect the issues
Orchestrate workflows
Script multi-phase, multi-agent pipelines for complex jobs.
Write a workflow: migrate all docs with per-phase validation
Load skills
Load specialized skills on demand: code review, web dev, docs writing…
Use the code-review skill to review this change
Long-running goals
Give one objective and the agent pushes it across rounds until done or blocked.
Goal: add bilingual pages to the whole help center
Plan & self-check
Lay out a plan before acting, then verify its own work with tests and builds.
List a change plan first; I will confirm before you start
Approvals & sandbox
Sensitive actions need your approval; sandbox and permission presets hold the boundary.
(no prompt needed — it is built in)
Advanced: it extends itself
With dynamic plugins the agent can mount new tools and UI blocks at runtime — "everything is a plugin" is not a slogan, it is how the harness actually works.
Dynamic plugins
Mount new capabilities at runtime, no source edits, no restart
Swappable models
DeepSeek, other providers, custom compatible endpoints
Controllable bounds
Approval policy, workspace allowlist, audit logs
Tempted? Up and running in three minutes
Follow the foolproof path from zero to your first finished task.
Take the foolproof path