Episode 40 | Why Your AI Agents Go Off the Rails — and the Harness That Saves Them | Ankur Bhatt
Ankur Bhatt — Head of AI at Service Titan — joins Amy and Meg to explain why most AI agent initiatives die between demo and production, and what to do about it. Ankur has spent the last two years building production agents that handle high-stakes work like tax notices and payroll compliance, and he's published one of the most useful practitioner guides on the topic anywhere. The answer, he argues, isn't a better model — it's something called harness engineering. He breaks down why agents have "the cognitive ability of a PhD with the attention span of a two-year-old," the three failure modes that sink most deployments, and the six principles that turn probabilistic AI into reliable enterprise software. Plus: why writing code is no longer the bottleneck, why your next product probably shouldn't have a UI at all, and a Leadership Corner on managing peer egos when you're the most senior woman in the room.
⏰ TIMESTAMPS:
00:00 Cognitive ability of a PhD, attention span of a two-year-old
00:17 Meet Ankur Bhatt: VP AI @ Rippling, Head of AI @ Service Titan
01:22 From SAP/SuccessFactors to startup speed
03:25 What customers actually want from AI right now
06:32 The demo trap: six-day demo, six engineers, three months of fixes
08:23 What "harness engineering" actually means 09:47 Why architecture matters more, not less, in the agent era
13:01 Where the term "harness" came from (the Manus story)
15:47 Three failure modes: compound error, context overload, specification vacuum
19:16 Why agents are like ADHD partners — the executive-function problem
21:11 The six principles of harness engineering
24:01 The Montessori analogy: maps, stations, and skills
27:43 Why specs and PRDs matter more now, not less (planning mode)
28:55 Skills vs. hooks: what goes where
30:57 Building a skills marketplace inside your organization
35:45 The 10–20% problem: scaling individual productivity to a team
39:40 The new bottleneck has moved upstream
42:52 From features to agent experiences (the Karpathy home-control example)
45:22 The two layers of B2B agent design every leader misses
48:46 Leadership Corner: lonely at the top, surrounded by egos
49:26 Meg's "trust council" reframe
53:26 Where to focus your emotional energy (hint: not on changing your peers)
55:31 Managing egos as a core executive skill
🔑 KEY INSIGHTS:
-Why your AI agent goes off the rails: compound error, context overload, and specification vacuum — and how to design around all three
-The six principles of harness engineering, in order — starting with "give agents maps, not manuals"
-Skills vs. hooks: how to encode domain knowledge and enforce quality without overloading the model
-Why "spec before code" matters more in the agent era than it did in the human-engineer era
-The new SDLC: when writing code stops being the bottleneck, what becomes the bottleneck instead
-Why continuing to build point-and-click UIs may already be irrelevant — and what an "agent experience" looks like in B2B
-Leadership Corner: why peer loneliness usually isn't a peer problem, and how to build a trust council instead
📚 RESOURCES:
Ankur's article: Agentic Engineering — Why the Harness Matters More Than the Model: https://www.linkedin.com/pulse/agentic-engineering-why-harness-matters-more-than-model-ankur-bhatt-fyjwe/
Daniel Kahneman, Thinking Fast and Slow
Andrej Karpathy on the No Priors podcast (the home-control agent example)
Anthropic's "progressive disclosure" approach to skills
Rippling: https://www.rippling.com
ServiceTitan: https://www.servicetitan.com
🤝 CONNECT:
Ankur Bhatt: https://www.linkedin.com/in/ankurbhatt77/ Instagram: https://www.instagram.com/megandamyshow/ LinkedIn: https://www.linkedin.com/company/the-meg-amy-show
#AIAgents #HarnessEngineering #AI #AITransformation #FutureOfWork #SoftwareDevelopment #Leadership #MegAndAmyShow
