Track 09 · Systems & unit costLesson 02 of 11
Agents where people already work
Slack, Teams and the tools of the day, not a new tab.
- Lesson library · reference
- 30 min read
- 60 min hands-on
- Pro
What you'll be able to do
01- You can explain why agents embedded in Teams, Slack, email or a CRM sidebar get adopted, and what the embedding costs in engineering.
- You can choose between a Teams SDK or Microsoft 365 Agents SDK app, a Copilot declarative agent, a Copilot Studio agent and a Slack Bolt app, and defend the choice.
- You can implement an ingress that verifies Slack signatures, acks inside three seconds, de-duplicates retries and replies asynchronously.
- You can map a chat user to an enterprise identity and enforce a deterministic rule about where a reply may be posted.
- You can render one reply as Block Kit and as an Adaptive Card with accessible fallbacks.
Before you start
- Agents started by events and schedules: webhooks, delivery guarantees, idempotency.
- Approval gates and bounded autonomy: gates before you put one in a channel.
- SSO and provisioning for agents: SCIM and IdP federation.
- Retrieval that respects who can see what: the requester's permissions decide what is retrieved.
- Agents inside CRM, ERP and ITSM: the systems behind the answers.
The short version
An embedded agent is a backend service that receives events from a collaboration platform (a mention, a DM, a button click), works out who the human is in the enterprise directory, does the work asynchronously, and posts a reply in the platform's UI format to a place where only the right people can see it. The platform gives you distribution and habit. In return it imposes a webhook contract (signatures, a three-second ack, retries), rate limits, a card format per surface, and an audience you do not control: a thread reply in a busy channel is a broadcast.
- Put the agent where the work happens: chat for coordination, the CRM or ITSM workspace for record-centric work, email for external and asynchronous flows.
- Verify every request, ack fast, process on a queue, de-duplicate on event ID.
- Map the chat user to an enterprise identity through a provisioned, immutable key. Unmapped means no answer.
- Decide placement in code from data labels and channel facts, not from the model.
- Keep one surface-neutral reply and render it per surface, with plain-text fallbacks.
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI