Synth AI
WorkshopBlogDocsStackSign in
WorkshopBlogDocsStackSign in
Changelog

Stack sidecar monitor — a separate stream for long agent runs

Goal mode in Stack now defaults to a Sidecar events feed: a monitor agent that narrates progress, audits done-claims, and steers the worker — instead of reading every tool call.

Wednesday, July 1, 20261 min read
stackagentsmonitor

TL;DR

  • /goal mode in Stack defaults to a Sidecar events feed — a higher-level stream, not the worker's raw tool tape.
  • A monitor agent runs 1:1 with your worker: it narrates meaningful progress, steers when the worker is stuck, and audits completion before flipping a goal to done.
  • Progress is a typed signal (stack_monitor_goal_status) rendered as a headline strip + milestone timeline, not a wall of checkpoint noise.
  • Quiet passes leave a dim check-in row so the feed proves the monitor is watching without inventing progress.
  • Available now on Stack dev / nightly.

What Shipped

Long agent runs turn worker output into thinking traces you skim. Stack's sidecar monitor gives you the stream you actually want: what the worker is doing, what milestone landed, and what went wrong.

In /goal mode the default view is Sidecar events. The worker's transcript is still there for debugging (a for the agent tape, t for the monitor's own thread), but it's no longer the primary surface.

The monitor posts updates through a typed tool, stack_monitor_goal_status, with a short headline, a one-line note, and an optional metric. The goal shutter renders those as a headline strip over a milestone timeline — so you track a run at a glance.

When the worker claims it's done, the monitor audits the claim against cited proof before marking the goal complete. A done-claim that doesn't clear the bar gets refuted in the feed as blocked or goal_failed, not parroted as success. On quiet passes — routine reads and listings with no real progress — the monitor leaves a dim check-in row (no change · reviewed N events) instead of faux updates.

Stack runs three agent roles on one runtime: the worker executes the goal, the monitor narrates and audits one run, and the gardener orients you across threads. Talk to the sidecar to follow a run; talk to the gardener for the portfolio.

Scope

This ships per one goal run: a separate higher-level stream, mid-flight progress, steering, and audited completion. Multi-goal portfolio rollups, ETA / progress-per-hour, and pause-before-irreversible are next.

bash
# on dev / nightly Stack
stack
/goal <your objective>
# watch Sidecar events; press a for the worker tape, t for the monitor thread
←Back to Changelog

Ready to try Synth?

Get started with serverless RL training and prompt optimization.

Get StartedSchedule Demo
On this page
  • TL;DR
  • What Shipped
  • Scope
© 2026 SynthWorkshopChangelogDocsBook a Demo