Skip to content
The Observatory

Building & Agents · 2 min read

De-Risk the Agent, Not Just the Business Model

Running Lean by Ash Maurya argues that founders waste time building the wrong thing because they never ruthlessly surface their riskiest assumption first. That discipline translates directly to AI-native companies, where the riskiest assumption is often not the market but the agent itself.

Featuring · studied & reframed for AI-native builders“Running Lean” — Ash Maurya

Find the Riskiest Assumption in Your Stack

Maurya's lean canvas forces you to name what must be true for your business to work, then attack the scariest item first. For a founder shipping with agents today, that item is usually trust: will the agent actually do the task reliably enough that a customer pays for the outcome? Before you run a single customer interview, build the evaluation harness that answers that question honestly. If you cannot measure the agent's failure modes, you cannot de-risk them, and you are skipping the hardest row on your canvas.

Problem Interviews Work on Workflows Too

Maurya's problem interview — talk to real people before you build — translates here into scoping your workflow before you grant it agency. Most founders reach for autonomous agents when a fixed-path workflow would answer the customer's actual problem cleanly. Interview the person whose job the workflow touches. You will discover that the branching they need is narrow and predictable, which means a workflow with clear tools and structured outputs beats a loosely prompted agent every time.

Solution Interviews Test the Human Handoff

When Maurya moves to solution interviews, he is testing whether the proposed mechanism actually solves the problem. In an AI-native product, the mechanism always includes a human judgment layer: where does the agent stop and a person decide? That handoff is the solution detail customers react to most strongly. Show them where accountability lives. If they cannot see it, they will not trust the product, and no amount of capability fixes a trust gap discovered after launch.

Iterate on Evals the Way Lean Iterates on Metrics

Running Lean treats one key metric as the north star at each stage and changes the plan when the number tells the truth. Treat your evaluation harness the same way. Pick the one failure mode that would kill the product — hallucinated output, missed edge case, wrong tool call — write red cases for it, and let those cases drive every design decision until the number moves. A good eval is a lean canvas row you can actually score.

Try the Desktop OS

Explore Atlanta’s startup scene, live.

Everything you just read lives inside our Desktop OS — an interactive workspace that opens right in your browser. Its Atlanta desk maps the city’s startup scene: who’s building, what’s happening, and where. Open it and take a look.

Open the Desktop OS →