User
Gives Cal a task in natural language.
Agent-to-agent orchestration
The Delegation Loop
Cal keeps the context, continuity, and decision loop. Codex works as a focused executor in its own thread. Questions return through the same visible Strand in the Cal UI for parallel sessions, then the work resumes where it stopped.
Cal owns the brain: context, memory, orchestration, and continuity.
Codex owns the run: focused execution inside a swappable worker thread.
One task, one visible Strand, one continuous executor thread
Cal orchestrates
- Prepares the relevant context
- Writes the task note
- Delegates the run
Codex Strand
The visible Cal UI for the delegated task and its complete timeline.
Working
Pending feedback
Ready
Background execution loop
The Strand stays visible while execution, analysis, and continuation happen behind it.
Codex executes
Works in its own durable thread.
- Reads, builds, tests, and reports
- Can be resumed after feedback
Cal analyzes
Reads the full Codex response, not a truncated preview.
Open questions?
Task completed
Ready · Summary posted
Cal closes the loop and updates the visible Strand.
Pending feedback
Attention · Needs input
- Cal drafts answers
- Cal notifies the user
Feedback policy
Summary to User
The same Strand holds the task, feedback exchange, resumed work, and final result.