Agents in the Wild: Production Agent Lab

09:00 – 10:15

See why laptop agents die in production

the demo-to-production gap, state and secrets audit, the plan

Takeaway Your agent's hidden assumptions listed: every path, key and in-memory state that will not survive a deployment.

10:30 – 11:45

Containerise the agent

images for agent workloads, configuration, secrets at runtime

Takeaway The agent in a container, configured from outside, with a planted secret kept out of the image and proved absent.

12:00 – 13:00

Deploy it as a service

the agent behind an API, health checks, versioned prompts and tools

Takeaway Your agent reachable as a deployed service with a health check that means something, and a version you can name.

14:00 – 15:15

Persist the state that matters

task state in a real store, resumability, what not to persist

Takeaway A long task killed halfway and resumed exactly where it stopped, because its state lives in a database, not a variable.

15:30 – 17:00

Handle the failures that will happen

tool outages, model errors, timeouts and compensation

Takeaway A tool taken down mid-task on purpose: the agent degrading, retrying and completing without corrupting anything.

09:00 – 10:15

Trace every run

run traces, step attribution, token and cost per task

Takeaway Twenty production-shaped runs on one dashboard: the expensive one found and explained in under a minute.

10:30 – 11:45

Budget and bound the fleet

concurrency and queues, rate limits, cost ceilings per task

Takeaway Fifty tasks submitted at once and handled on your terms: queued, bounded and each under its own cost ceiling.

12:00 – 13:00

Keep humans in the loop at scale

approval queues, escalation paths, audit trails

Takeaway Approvals arriving in one queue a human can actually work, with every decision recorded for the audit that will come.

14:00 – 15:15

Rehearse the bad day

incident drill, the kill switch, stopping without losing work

Takeaway A live incident drill: the agent misbehaving on cue, caught by its traces, and stopped with the kill switch you tested.

15:30 – 17:00

Write the runbook and hand it over

what to monitor, when to intervene, handover review

Takeaway A one-page runbook another engineer could operate your agent from, reviewed by the room as the final exercise.