Workflow guides
Run, isolate, coordinate, verify, and deliver agent work safely in Tejus.
Tejus workflows are built around one rule: every body of work needs a clear owner and a verifiable result. The tools—threads, worktrees, forks, handoffs, the in-app browser, automations, and agent-to-agent coordination—are useful only when they preserve that boundary.
Start with the smallest workflow
| Need | Start here |
|---|---|
| One focused change | One thread in a known checkout |
| Independent work in the same repository | Separate worktree-backed threads |
| Another path from the same conversation | Thread fork |
| Another provider continuing the same objective | Provider handoff |
| A defect whose cause is still unknown | Debug mode |
| Visual or web behavior verification | Browser verification |
| Review and delivery through GitHub | Pull requests |
| Recurring work | Automations |
| One Tejus agent coordinating other Tejus threads | Agent Gateway |
| Another local application creating restricted threads | External MCP |
Do not begin with maximum parallelism. First learn the complete loop: objective, execution, supervision, verification, review, and delivery.
Core workflow
Begin from a known state
Confirm the project, branch, existing changes, provider, model, and whether the thread works in the local checkout or a worktree before sending work.
Give one owner a bounded objective
State the outcome, scope, constraints, and verification. Split work when those cannot be expressed clearly.
Supervise the work
Watch plans, tool activity, approvals, questions, background tasks, and the growing diff. Correct wrong assumptions early.
Verify independently
Inspect every changed file and run the relevant checks. Use the in-app browser when visible behavior matters, and read any Security review findings reported after the turn.
Preserve and deliver intentionally
Commit only reviewed changes. Push, open a pull request, or archive the thread only after its ownership and state are understood.
Guides
Parallel threads, handoffs, automations, and orchestrators can increase throughput. They also increase the number of assumptions, diffs, and failure states you must verify. Keep one clearly identified integration owner responsible for the final result.