On November 3, founder and Late Checkout CEO Greg Isenberg (@gregisenberg) posted a long thread on X saying his team has 50+ people in 12 countries, yet he "talks to my agents more than my human teammates." He listed an agent-driven stack covering ideation, design, build, user testing, bug fixing and data review. Within hours it drew about 1,050 likes, 2,300 bookmarks and 183 replies. This is a personal workflow and opinion, not verified productivity data.
Key Takeaways
- ✓Engagement: about 1,050 likes, 2,300 bookmarks, 183 replies and 75K views within hours (per the X API); bookmarks far exceed likes, so readers are saving it as a checklist
- ✓Coverage: ideation, design, build, user testing, bug fixing and data review are all handed to agents; humans handle "creative feedback, relationships and telling the agents what they got wrong"
- ✓Build: pingdotgg/t3code as the control surface and stablyai/orca to run parallel agents, each in its own git worktree
- ✓Quality control: every output is scored 1-10 by an LLM judge and anything under 8 goes back; every agent mistake becomes a new skill
- ✓All of this is self-reported; no verifiable productivity or cost figures were published
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
By 2026 agents can take on more end-to-end work, but concrete examples of how a small team actually chains them into a production line are still scarce. Greg Isenberg (@gregisenberg, CEO of Late Checkout and host of the Startup Ideas podcast) drew heavy bookmarking because his thread lays out a full checklist from idea to launch. His opening line, "running a company nowadays feels kinda dystopian," also sparked debate about what human teammates are for.
How it works
As he describes it, ideas go to an agent by voice note, mattpocock/skills /grill-me makes the agent interrogate the idea, and a spec lands in Linear via the Linear MCP. In design, Claude Code with Paper MCP produces four versions of each screen and merges the best parts. In the build, t3code is the control surface and orca runs a fleet of agents in parallel, each in its own git worktree, with LSP plugins and a security hook. For testing, the Lightpanda headless browser runs agents through every flow as impatient, confused and mobile personas. When Sentry MCP catches an error, an agent reads the stack trace and opens a fix PR; a human just approves.
Benchmarks
There are no public benchmarks. Figures in the thread such as "a fix PR in under 10 minutes" or "tokens go 20-95% further" are the author's own claims or third-party tool marketing, not independently verified. Only the engagement numbers on the post can be checked.
Getting started
The transferable part is the method, not the tool list: one agent per platform, a cheap model for triage and a strong model for anything that ships, an LLM judge as a quality gate, a five-line nightly log per agent, and turning every mistake into a skill. Pilot one stage first, such as bug fixing or a daily funnel review, and keep human approval for anything touching databases or payments.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.