Why Your Agents Fail. Lessons from Harness Engineering
You can ship an agent that demos beautifully. Production is a different story.
Most teams in 2026 are tuning prompts and swapping models. They're spending their time on the wrong layer. The hard part isn't the model. The frontier models are good enough. The hard part is the engineering around it: the state machines, the validators, the manifests, the gates. The harness.
This is a free three-hour masterclass on the harness layer, taught with running code from real production systems. By the end you'll have a vocabulary for why agents fail, a checklist of patterns that fix them, and one tactical thing to do at work on Monday morning.
This event is made possible by Rasa, CDTM and ZenML.
It's also Episode 0 of Why Agents Fail, a free online course launching this summer at Rasa University and profrod.ai.
What you'll walk out with
A vocabulary for talking about agent failure, "harness failures" vs "model failures," and why the distinction changes how you debug
A pattern library of harness fixes that reappear across voice, text, and document agents
Live demos of agents failing in production-realistic ways, then being fixed without changing the model, running Python code, all of it open-sourced
A specific Monday-morning action for whatever you're building right now
The course foundation. Saturday is Episode 0. You'll know whether the rest of the course is for you
The agenda
Three hours, structured around real demonstrations. Two ten-minute breaks built in.
14:00 — 14:20 · The thesis Why most agent failures in production aren't model failures. The data points anchoring the field in 2026, and why the leaderboards you read are partially noise.
14:20 — 15:10 · Block 1 The first set of failure patterns, demonstrated end-to-end. Each one comes with a runnable Python script (open-sourced after the session) so you can reproduce it.
15:10 — 15:20 · Break
15:20 — 16:10 · Block 2 The second set of patterns. The ones that show up across very different agent shapes (voice, chat, document workflows) but turn out to share a root cause and a fix.
16:10 — 16:20 · Break
16:20 — 16:50 · Block 3 Synthesis. The harness lens applied to your current project. Cost engineering, observability, evaluation, deployment. The parts the demo videos always skip.
16:50 — 17:00 · Q&A
17:00 · Social & Networking