Skip to main content

Watching sub-agents

A delegated task is not a black box: caesar watch opens a window onto what is happening, next to the conversation or terminal that launched the delegation.

caesar watch                 # all running tasks, redrawn frame
caesar watch t_a1b2 t_c3d4   # only these
caesar watch --once          # one frame, then exit
caesar watch --json          # NDJSON of the events, several tasks merged
▞▚ caesar · watch   1 active · max_parallel 4                             17:21:20

● t_efb5914d codex        —            25s  inplace · write
  Write three files a.txt, b.txt and c.txt, then run 'sleep 8 && ls -1'…
  ▸ shell /bin/zsh -lc 'sleep 8 && ls -1' — 3s
  ~ 3 file(s)  ·  11 event(s)

q or Ctrl-C to quit — watching modifies nothing.

No daemon is needed: the engine writes events.jsonl during execution and publishes task state through atomic writes. caesar watch only reads what another process writes — the same property that makes caesar cancel and the sharing of max_parallel work.

Four deliberate design choices​

  • A tool appears as soon as it starts, not at its completion. That is the whole difference between seeing a three-minute npm install set off and discovering it at minute three.
  • Silence is displayed. A stuck task and a working task are indistinguishable without it; past thirty seconds without a single event, the view says so.
  • A pending question jumps ahead of everything else. A sub-agent waiting for an answer over the return channel looks exactly like a frozen sub-agent.
  • Finished tasks stay visible for a few minutes, with their report status: a task that disappears the moment it finishes is a task whose ending you will never know.

Outside a terminal (redirection, | tee, script), no redraw and no ANSI sequences: one line per event, and --json yields usable NDJSON.

What each agent lets you see​

What you can observe during execution depends on what each CLI narrates, and that varies a lot:

AgentDuring execution
codexstart and end of every command, modified files, its progress reports
claudetools, results, text, and an in-progress thinking signal
opencodetools (only once finished — its stream does not announce their start), text
antigravityits text as it streams, its errors; its tool calls are not yet translated
copilottext, session errors; its tool calls remain unverified for lack of available quota

These translations are written from real captures and replayed by tests. Where a shape could not be observed, the adapter says so in plain words rather than guessing.

Interactive view — not for scripting

caesar watch without --once redraws and never terminates on its own; it is meant for a human at a terminal. For a single snapshot in a script or an automated flow, use caesar watch --once or caesar ps instead.

Other useful subcommands while following delegations: caesar ps (running and recent tasks), caesar logs <id> [--raw] [--follow], and caesar cancel <id>. See the CLI reference for the full list.

Next steps​