Task Checklist
Agents keep a visible to-do list as they work, checkpointed with the run and rendered live in the run log.
An agent working through something long keeps a task list. You see it fill in and tick off as the run goes, in the run log, live.
The list is written by the agent itself through a native update_tasks tool. There is nothing to enable — every agent carries it.
| Status | Meaning |
|---|---|
pending | Not started |
in_progress | Being worked on now |
completed | Done |
cancelled | Abandoned, deliberately |
Why it is part of the engine
It replaces the @modelcontextprotocol/server-sequential-thinking MCP server, which took a thought, echoed it back and kept a list in its own process memory. Nothing reached the journal, nothing rendered in the UI, and nothing survived a resume.
The task list is the opposite by construction:
- The engine owns the state. It lives on the run and is checkpointed with it, so a run that pauses at a human gate or is interrupted comes back with its list intact.
- Every change is journaled as a structural event, so it survives the event cap that trims chattier events on a long run — a task list is the run's shape.
- The run log renders it as a live checklist rather than as tool-call noise.
It is not an MCP server. It is imported directly by the engine, so there is no subprocess and nothing to install.
Whole-list writes
The model submits the complete list on every call, rather than patching individual entries. That looks wasteful and buys three things:
- One flat schema with no type unions — the exact shape that broke Gemini's single-valued schema type.
- No task ids for the model to fumble.
- The engine's identical-arguments loop guard becomes effective: resubmitting the same list verbatim is a no-op and gets blocked. The old server's ever-changing thought counter defeated that guard by construction.
sequentialthinking still works
Agents and imported bundles that name the old tool keep working — the dispatcher routes the alias to the native handler. Update them when convenient; the alias is one release of grace, not a permanent surface.
Synapse AI
Synapse AI is the product's assistant: ask it about your workspace, have it run or answer things with a card you press, change what exists through a change you approve, or describe a job and approve the plan it builds.
Scheduling
Schedules let you run agents or orchestrations automatically on a time-based trigger — either on a fixed interval or a cron expression.