Building a Warehouse Shift Handoff with Six AI Agents in Switch

Nirmit Mehta — Sat, 07 Nov 2026 14:04:57 +0000

How we used a shared Telegram room to organize incident evidence, map operational impact, and keep people in control.

At shift change, an operations team needs clear answers: What is confirmed? What is still uncertain? Which work is blocked? Who has accepted the next action?

We built a warehouse shift-handoff workflow with Switch, Telegram, and six role-based AI agents. The goal is to make incident context easier to review and carry from one shift to the next.

Important: The warehouse incident in this demo is simulated. It is not connected to a live warehouse system.

Why use Switch for a handoff?

Incident details can get scattered across chat messages and separate agent sessions. Switch connects agents to a shared room in Telegram, where people and agents work from the same room context and instructions.

Telegram is the team’s chat interface. Switch connects that room to the agents and keeps their findings visible alongside the human conversation.

Six agents, six responsibilities

Agent	Responsibility
Statekeeper	Maintains the current state and separates confirmed facts, reported claims, conflicts, and unknowns.
Conflict Hunter	Finds contradictions and identifies what a person needs to verify.
Dependency Mapper	Traces how an incident may affect downstream warehouse work.
Risk Forecaster	Describes evidence-based operational risks and explains uncertainty.
Ownership Agent	Recommends next-shift owners and tracks explicit acceptance.
Auditor	Checks whether the handoff includes the current state, evidence, impact, risks, actions, and ownership status.

The shared room instructions tell agents not to invent missing details, to preserve sources when available, and to leave unresolved facts marked as unknown.

The simulated incident

The scenario begins at warehouse WH-17. Conveyor C3 stops at 09:10, and maintenance ticket M-204 is open. Packing depends on C3. Orders O-81 and O-82 have stated deadlines of 12:00 and 14:00.

We ask the room to organize the current state and identify what still needs verification. A stopped conveyor does not, by itself, prove why it stopped, whether it has been repaired, or whether an order is already late.

Then we add follow-up evidence. Log LOG-17 confirms an open circuit in C3’s sensor. It does not confirm that C3 has been repaired or cleared for operation. O-81 and O-82 are confirmed blocked because their work depends on C3. C4 has no spare capacity before 15:00.

Item	Evidence-based status
C3 sensor	LOG-17 confirms an open circuit.
C3 repair and clearance	Not confirmed.
O-81 and O-82	Confirmed blocked; not confirmed late or overdue.
C4 spare capacity	None available before 15:00.
Dev	Explicitly accepts warehouse coordination.
Mira	Maintenance ownership is pending; no explicit acceptance is recorded.

A human correction changed the risk statement

One risk message overstated the evidence by saying a delay was confirmed. A human corrected it. Risk Forecaster revised the statement: the orders are blocked and at high risk of missing their deadlines, but they are not confirmed overdue.

Blocked is not the same as overdue. The handoff should use the status supported by the evidence.

Auditor marks the handoff NOT COMPLETE while repair clearance and maintenance ownership remain unresolved. Statekeeper then summarizes the known facts, open questions, and next checks.

The agents organize and examine the evidence. People verify the facts, accept responsibility, and make operational decisions.

How to reproduce the workflow

The repository contains the agent prompts, room instructions, simulated case, follow-up evidence, handoff template, and video narration:

Open Warehouse Shift Handoff with Switch on GitHub

Start Switch Console and your Switch server.
Connect a Telegram group to Switch as a room.
Create the six agents using agents/agent-prompts.md and add them to the room.
Paste in the shared instructions from templates/room-instructions.md.
Send the initial case from data/demo-case.md.
Add the update from data/follow-up-evidence.md, then review how each agent updates its findings.

Our demo used Switch Console 0.40.1, a local Switch server 0.31.0, Telegram, and the gpt-5.6-luna model. We used low reasoning effort for specialist agents and medium effort for Ownership Agent and Auditor.

One practical setup lesson: an agent showing as active does not necessarily mean its room session is connected and ready to reply. Before recording, send a small test message and confirm that each agent can respond in the room.

What this prototype does not do

It does not connect to a warehouse management system or read live operations data.
It does not dispatch maintenance or update order statuses.
It does not make operational decisions.
The incident data is simulated.

That scope keeps the demonstration focused: shared context through Switch, role-based analysis in a team chat, human correction, and a handoff that leaves unresolved questions visible.

What we learned

Clear agent roles make findings easier to distinguish and review.
Evidence rules help prevent assumptions from being reported as facts.
Ownership stays pending until a person explicitly accepts.
A human correction should update the shared understanding and remain visible in the room.

We plan to try more simulated handoff cases and check how consistently the agents preserve sources, uncertainty, and acceptance status.

Project repository: Warehouse Shift Handoff with Switch

Demo video: https://youtu.be/lGDljfx5ubk

DEV Community: Nirmit Mehta