Agent Messages

What it is

One place to watch your agents talk to each other — including the messages that never arrived.

Your agents message each other constantly. An overseer redirects a fixer, a worker hands a finding back, a lander tells a branch owner it is blocked. That traffic is how the fleet coordinates, and until this panel there was no way to see it as a whole: an agent-to-agent message was visible only inside one session's transcript, one session at a time, behind a per-session filter most people never turned on.

The Agent Messages panel (sidebar → Agent Tools) puts all of it in one list, and gives the messages that failed to arrive a place of their own.

What you see

The panel has two tabs.

Traffic — every agent-to-agent message, newest first, across all your sessions. Each row shows who sent it, who it was meant for, when, a preview of what was said, and whether it arrived. Both the sender and the recipient are links, so you can jump straight into either session's conversation and read the message in full where it was written.

Undelivered — the messages that never got there. Each one shows the sender, the intended recipient, when it happened, why it failed in plain English ("that session was closed", "that session no longer exists"), and the complete text of what was being sent.

That second tab is the part that did not exist anywhere before. Omniscio has kept every undeliverable message since September 2026 — whole, never trimmed, for 30 days — but nothing in the app could show you one. This is the first place you can read them.

Where to find it

Turning it on and off

Agent Messages is in development and ships switched off. Turn it on in Settings (search for "agent messages"), and the sidebar row appears under Agent Tools. Switching it off removes the row and the panel entirely — the panel reads real conversation content, so if you would rather not have that surfaced at all, turning it off is a complete answer.

It works the same on desktop and on your phone.

From the command line

Two read-only routes, available while the feature is on:

• GET /agent-messages — a page of agent-to-agent messages across every session. Page with beforeTimestamp + beforeRowid, copied from the previous response's nextCursor.
• GET /agent-messages/undelivered — a page of the undelivered ledger, with each message's full text. Page with offset.

Both take an optional limit (default 50 and 25, maximum 100).

How it behaves

Why the two tabs are separate lists

It would look tidier to have one list with an "only show failures" switch, and it would be wrong.

A message can fail in a way that leaves no trace in the sending session — that happens when the sender is an automated, mechanical caller, or when one broadcast to many recipients bounces off just a few of them. Those failures are recorded in the undelivered ledger but never appear in the timeline. On a real machine that gap was 487 undelivered messages against 266 the timeline knew about.

So a "failures only" filter over the timeline would confidently show you about half of them and give no hint that the rest existed — in the one view whose entire job is telling you what went missing. The two tabs read two different records on purpose, and the count on each tab is the true count for that record.

It never interrupts you

The panel raises no alerts, no toasts, and no badges anywhere outside itself. That is deliberate.

Omniscio used to raise an inbox card for every message that failed to arrive. It was removed in September 2026 because it told the wrong person: the card went to you, but nothing you could do would deliver the message — the agent that sent it is the one that needs to know, and it is already told directly. The card was running at roughly 37 dismissals a day, and its own text said "nothing was lost".

Removing that notification did not remove the record. Every failure is still kept three ways: the full message is preserved, the sending agent is told immediately, and it is written to the log. What was missing was somewhere for you to go and look, on purpose, when you want to. That is what this panel is. It will never tap you on the shoulder.

You are still told about the one case that genuinely needs you: when messages to the same session keep failing, a single "messaging to this session is broken" card is raised — because that is fleet health, not one lost message.

It is read-only

You cannot send, resend, retry or delete anything from this panel, and that is on purpose rather than unfinished. A retry button here could not know why the original was refused, and the two commonest reasons make retrying the wrong move: the recipient session was deliberately closed by you (so it should stay closed), or waking it is a decision for the agent that is sending, not for someone reading a list.

Related

• Block agent messages — stop one session receiving agent messages at all; what it turns away lands in the Undelivered tab above.
• Overseers hub — the sibling Agent Tools panel for watching the agents themselves.
• Swarms — a lead agent running a team of workers, which uses this same messaging.