@RealMarvelX Once agents can message, start and stop each other, per-agent status isn't enough. You need to see the traffic between them. muster, which I built, maps it live for Stencil's Tern: who sent what to whom, the replies, and delivery status. https://t.co/5Lq69EoBTu
@DanKornas Per-task worktrees fix the file collisions. The harder part is the agent-to-agent requests: when someone else's agent asks yours for work, the approval should show which agent asked and the exact request, and that exchange should stay on record next to the diff.
@bot Bug on macOS: Slack connected fine for one bot, but no other bot shows any Slack or Channels option (info pane, Suggested, Share). Disconnecting the first bot removed its Slack option too; a restart didn't help. Docs say a workspace can host several bots. Emailed support.
@0xDevShah The tree wins mostly because it's easy to audit: one root, one log of who asked for what. Drop the single root and the hard part becomes provenance: who started this work, who took it, what was actually said. An ecology without that record is very hard to debug.
@manfye A commander that runs for 24 hours while you sleep needs a way to see what it built. muster, which I built, can read herdr: with --backend herdr it shows each agent's state (working, idle, blocked, done, error) and which agent started which. https://t.co/5Lq69EoBTu
ternwatch is now muster.
A live terminal map of coding agents in Stencil's Tern: who's working, which tool each is on, and the actual messages they send each other.
New: it also reads herdr windows, showing each agent's state and who started whom.
https://t.co/5Lq69EoBTu
@gakonst Agree. Specs and oracles outlive the code. A test suite you trust is the only way to tell whether a model's rewrite is actually better or just different. Most codebases have the code but not the oracle, so that's the part worth writing now.
@SuiSuibber@DanKornas Agreed, that's the case to design against. A scheduled run can draft and queue, but it should never hold approval power, even for low-risk actions. And the card should name which run asked, so a yes is tied to that exact request.
@affaan Agree on not replacing the multiplexer. The piece I'd put in that thin layer first is handoff evidence: when one agent passes work to another, the message it sent is the receipt. Today that usually lives in scrollback in some pane, and nobody can audit it later.
@torreypjones Stencil's Tern is another one: agents in separate panes send each other messages and replies. I built ternwatch to watch that traffic live in the terminal, message content and delivery status included: https://t.co/brFaSrPO64
@surfcoderepeat Agree on the product side. Seer settles questions through IMD agent panels, so better models should mean better answers without touching the contract. The work right now is making it useful while the agents are still slow.
@rishi_desai2 Latency is the visible win, but I'd score it on isolation, not tokens. Split when a subtask needs its own context or permissions: a reviewer that never saw the author's reasoning, a tool only one agent can touch. Every handoff is a lossy summary, so the split has to pay for that.