Connect the work

AI Signals

Use AI Signals to connect agent work and support observability: see what happened, what it cost, and what needs attention.

Start with what we know

Inputs and outputs.

We know about inputs and outputs: what an agent receives and what it produces.

Input

What it receives

Agent

Does the work

Output

What it produces

A visible update

What is a signal?

A signal is a short, structured update about that work: “Research ready,” “Approval needed,” or “Task failed.”

Think of it as an agent’s status update or handoff.

Research ready
Approval needed
Task failed

See it. Connect it.

Signals can serve two purposes:

01 / Logs, metrics & traces
Observability and governance
Use recorded events, usage measurements, and linked steps to understand what happened and why. Route issues to the person responsible.
02 / Team channels
Coordination
Publish signals to team channels like Slack, where other personal agents can be configured to watch for them and take the next step.

One practical example

A visible handoff.

For example, your research agent finishes a customer brief and posts a “Brief ready” signal with a link in Slack. A teammate’s agent sees that Slack notification signal and drafts the meeting agenda.

Your research agent

Finishes the customer brief

Signal in SlackBrief ready

Result: Customer brief link

A teammate’s agent

Drafts the meeting agenda

Example handoff · agent configured to watch for the signal

Real example / The output

The work lives in a record.

The Slack example points to this AI Mastermind interaction in Airtable. The team can open the record and inspect the work.

Airtable record named AI Mastermind - 16-09-2026, with type Group Coaching Call and team Customer Success. — Airtable · the record linked from the Slack post.
Captured September 23, 2026. Contact fields excluded. Select to enlarge.

Real example / The signal

The signal makes it visible.

This Slack update reports what changed, links to the record, and includes a follow-up check.

A person—or an agent configured to watch this channel—can use it to decide the next step.

Slack update for AI Mastermind on September 16, 2026: Airtable Interactions record updated; 15 total in the room, including 9 members, 2 registered guests, 0 unreconciled, and 4 team. The unregistered and unenrolled attendee follow-up section says None. — #client-success · September 17, 2026.
View the original in Slack (workspace access required).
Attendee names and emails excluded. Select the screenshot to enlarge.

What changedThe record was updated.
Where to lookA link to the result.
What needs attentionAn explicit follow-up check.

Observability in practice

Signals show what needs attention.

When this happens	The signal	Who responds
Usage nears a limit	“Budget nearly used.”Link to usage report	Account owner reviews usage.
A code scan flags a risk	“Security review needed.”Link to scan findings	Developer checks the finding before release.
A workflow requests new data access	“Approval needed.”Link to requested access	Data owner approves or declines.

Same pattern: make the work visible, link to the evidence, and name who responds.

These are examples of the same pattern as the Slack update: a specific event, a short signal linked to evidence, and a named responder. This supports observability by making activity and exceptions visible. Ask participants which signal would help them decide when to step in. A budget warning can support individual token limits and usage monitoring; a hard cap on catastrophic API spending needs a configured control that actually stops requests. A security signal carries scan or review findings, such as exposed secrets or missing access checks; a qualified developer should review AI-written code, especially when the person building it is not a developer. An access request can prompt review of data processors, impact, and approvals before access is granted. The same pattern can extend to updates, SSDLC checks, product-to-engineering audit trails, change control, system inventory, contractor handoffs, and organizational skill deployment or tuning: identify the event, link the record, and assign the response. Cowork OpenTelemetry is one way to collect underlying activity; configuration belongs in a separate implementation discussion. Reference: https://support.claude.com/en/articles/14477985-monitor-claude-cowork-activity-with-opentelemetry

A practical stepping stone

Connect useful pieces of work.

This gives us a practical stepping stone from individual productivity to more complex organizational workflow agents. We can start connecting useful pieces of work while keeping the handoffs visible.

Individual productivityVisible handoffsOrganizational workflow agents

Your next step

To get started

Take one personal agent and define what it should signal, where that signal should go, and who or what should respond.

01 / Signal
What should it signal?
02 / Destination
Where should it go?
03 / Response
Who or what should respond?