Multi-agent coding: divide tasks, verify results and bring them together
Divide a larger task into clear responsibilities, retain member conversations and task activity, then deliver against actual checks.
Define a goal and acceptance checks
For example: “Add theme switching after reading the existing styles. Run the project checks, list changed files and document unresolved issues.” A concrete scope makes completion easier to assess.
Use /plan when you want to review the approach first. Plan mode supports reading and planning, not project edits. Explicitly approve the plan before moving to execution.
Keep responsibilities visible on the task board
- Split independently actionable parts, such as styles, component changes and test review.
- Use Agent Team to inspect member status, conversations, owners and dependencies.
- Clarify a member’s scope or stop its task as needed, then collect the actual outputs and check results.
- Several sessions for the same user can write to the same directory. Use separate directories or Git worktrees to avoid concurrent edits to the same files.
Skills, Workflow and persistent goals have different jobs
Skills reuse methods and conventions. Workflow arranges stages or parallel work with visible logs. Agent Team provides members and a shared task board. None guarantees successful completion.
A persistent goal requires explicit /goal activation. You can stop it and explicitly resume it; ordinary requests do not automatically create one. Model failures, execution limits and permissions can still interrupt progress, so inspect the actual state.
Review outputs and record the limits
Check the changed files, executed commands, check output and unresolved issues before accepting the result. Child and parent tasks remain subject to the selected permissions and model access.
Ternilo Work is planned: one persistent Docker container for multiple workspaces, sessions and Agents. Container management and scheduling under Work are not currently available.
Common questions
Does refreshing lose the Agent Team?
Saved conversations and parent-child relationships are retained. Whether a run continues depends on its execution service; saved history does not automatically restart an interrupted process.
Are more Agents always faster?
No. Independent tasks may benefit from delegation, while shared files, dependencies and model budgets can constrain parallel work.
Can a goal run continuously without explicit activation?
Persistent goals require explicit activation and still obey permissions, approvals and execution limits. Use the provided stop and resume controls.