Ship AI agents that survive production
One hour on the failure modes that only appear after the demo works.
Wednesday, 12 August 2026 at 18:00 · 60 min · free to attend live
A working session on the failure modes that only appear after the demo works: authority boundaries, duplicate effects, and recovering a run that stopped halfway.
Save your seat
Attending live is free. Register below — the stream will appear right on this page when we go live, and as a subscriber you keep free access to the recording.
You'll get the live link before we start. One email a month, unsubscribe anytime.
Starts in …
Who this is for
- Backend engineers putting an LLM behind a real workflow
- Tech leads who own an agent nobody fully trusts yet
- Anyone who has watched a retry quietly charge a customer twice
What you will leave with
- A written authority boundary: what the agent may decide alone, and what stops for approval
- The three checks that catch duplicate effects before they reach a provider
- A recovery path for a run that stopped halfway, without replaying the transcript
What we will cover
- Starts at 0:00Why demos lie
Clean inputs, a patient operator, and no history. Production has none of those.
- Starts at ~15 min inDrawing the authority boundary
Where approval belongs, and why prompt text is not an enforcement point.
- Starts at ~35 min inEffects that run twice
Idempotency keys, timeouts that lie, and reconciling against the provider ledger.
- Starts at ~50 min inLive questions
Your host
Jakub Jirák
Engineer, thinkdifferent.blog
Writes about running AI systems in production, and the operational details that decide whether they hold up once real traffic arrives.
Questions
Is it recorded?
Yes. Newsletter subscribers and anyone who has bought a course get the recording for free.
Do I need to prepare anything?
No. Bring a workflow you are nervous about — the examples work better against a real one.
Is attending live free?
Yes, attending live is free. Register and the stream appears on this page when we go live.