SpaceXAI engineer building Grok Bot just gave a free 55-minute masterclass on how to trust coding agents. Lauren Tan went from micromanaging one agent to waking up to 20 auto-merged PRs. Here’s the full journey: 0% → 00:00 - the trust curve: why you can’t scale from 1 agent... to 100 without verification 25% → 14:07 - how she built Pstack: every failure mode became a new skill 50% → 26:53 - verification first, then cloud agents, then automatic PR merges 75% → 42:43 - why the shortest path should be the best path for your agents 100% → 50:42 - how strict architecture lets PMs and designers ship code safely Trust → Verification → Skills → Evals → Cloud Agents → Hard Constraints → Autonomous Engineering The most important idea: Don’t tell agents to “be more careful” Build an environment where the easiest thing for them to do is also the correct thing. Bookmark it, watch the full talk, and start building your own agent workflow 👇show more

a 55 minute masterclass on trusting coding agents is worth more than most evals: trust is a review habit, not a score. the part people miss is that the agent has to show its work before you can decide whether to trust it.

Where is this sneaky Chinese girl from? Is she a Chinese spy?🪳🦗

20 merged prs overnight is wild

does she talk about what happens when the cheapest path for the agent is also wrong? that trust curve seems to assume verification costs stay low

waking up to 20 auto-merged PRs sounds incredible, trust really changes the game

trust doesn’t scale by hoping the agent gets smarter. it scales by turning every failure into a constraint, until the safest path is also the fastest one.

the trust curve is really interesting

good

Facts!

that's interesting idea bro

The line about making the correct path the easiest path is the part worth keeping. Most agent failures I have seen were not judgment problems, they were friction problems, where the right move cost three extra tool calls and the wrong one cost none. You can ask an agent to be careful, or you can price the shortcut out of the room.

this talk deserve to be saved

the jump from 1 agent to 20 auto-merged PRs is the scary one. i only got there after reading every diff myself for a month.

Grok became useful in my team. I'm using Grok more and more as it's faster and cheaper.

The trust curve is really an operating-system problem. More agents only help when verification, merge policy, and rollback scale with them. Otherwise the human remains the bottleneck, just with a larger queue of things to approve.

this girl did a great job actually

Trust is the actual product, and it never came from one landing page for my MVP. It came from a sequence: content earns attention, a profile earns credibility, a real result lands before I ask for a cent. Skip a step and the whole chain reads as hollow.

‼️