Workflows & choosing a model
How you organise your work around an agent has evolved fast — from micromanaging every diff to leaving a Ralph Loop running overnight. Here are the six workflows, how to pick one, and how to compare models honestly.
01The evolution of workflows
A workflow is how you organise your activity around the agent. It’s exactly what Karpathy’s tweet was about, and it has shifted with the two mindsets. Think of six levels of trust — loosely mirroring yesterday’s eight stages.
The 2025 mindset — three levels of trust
1 · Micromanagement
Very specific instructions in agents.md, approve every change, frequently stop, rewrite,
reset and re-run. Careful and hands-on — where most people (and the instructor) began.
2 · Plan → execute → review → test
Use plan mode: the agent writes a to-do plan; you agree, then switch to execution. Build in phases — for each phase: build, code-review (stop it making a mountain out of a molehill), test, review results, mark complete, move on. A bit more rope, still on top of things.
3 · Spec-driven development (“trust but verify”)
Specify precisely what’s needed — sometimes in a formal spec language — then let it run. You verify at the end rather than reviewing every step. A real step back in oversight.
The 2026 mindset — three more levels
4 · YOLO
No approvals — the agent does whatever it wants, no permissions. “Trust but verify” taken further: set it going, go have dinner, come back and see what it did. Around most of 2025 this was hobby-only; now people use it for real.
5 · Ralph Loops
Named after Ralph Wiggum (naive, optimistic) and invented by Geoffrey Huntley. An agent already loops internally; a Ralph Loop wraps that whole loop in a bigger loop: run it, test “is this good enough?”, generate feedback, add it to the objectives, and run again — up to, say, 10 big loops. Leave it overnight; come back to a lot done.
6 · Multi-agent / swarms
Many agents with different roles — testing, feedback, manager and worker agents in a hierarchy. Spawned in numbers and orchestrated. This is the frontier in 2026.
02Which approach is right?
The classic answer: no single approach is “right” — different approaches suit different tasks. Split it in two:
Stay hands-on (levels 1–3)
Enterprise software, commercial SaaS, large codebases, or highly innovative code (e.g. new MCP servers models handle poorly). Precision and oversight matter most.
Let it run (levels 4–6)
MVPs, prototypes, pilots, brand-new empty directories, lots of boilerplate (HTML, a React app, CRUD backends), and some risk appetite. Great fit for YOLO and Ralph Loops.
In practice the instructor lives mostly in the top category — mission-critical, often highly innovative work like MCP integrations, where current LLMs write non-idiomatic code because the pattern is too new. But yesterday’s game was the other category: essentially YOLO to start, then a Ralph Loop for the fancier version. The course leans toward mission-critical, large-codebase, real enterprise patterns — with a solid nod to the 2026 techniques.
Your job is to deliver code that’s proven to work. “The LLM wrote it” is no excuse. Use LLMs freely — they let you do far more — but it’s still your responsibility to check, validate, and pick the right approach for the task. Take accountability for what you ship, whether or not an AI helped you get there.
03Beyond the hype
A sobering, honest note from a genuine enthusiast. There is unwarranted hype out there.
Order-of-magnitude faster
Boilerplate-heavy work with a clear target — a React front end with lots of components. Minutes instead of a day for a strong front-end dev; days-into-minutes for greenfield projects.
Only incrementally faster — or slower
Highly innovative work on large codebases. Sometimes it has actually slowed things down — a subtle, unexpected mistake that had to be found and rewritten.
Net-net: LLMs are a genuine multiplier — but not a flat 10×. How much depends heavily on the project. Our job is to learn the tools, use them, and set the record straight — for both the over-hyped and the newly anti-AI crowd — explaining where they shine and where they still need to grow.
04Comparing models — artificialanalysis.ai
If you keep one bookmark, make it artificialanalysis.ai. It compares models on intelligence, speed and price, dimensioned every way you’d want.
The top models cluster around a late-2025 inflection point when coding agents took a step-change in reliability. Some sceptics still carry scar tissue from the pre-November era — it’s worth encouraging them to try again, because so much has changed. On the same site you’ll also find a frontier-intelligence-over-time chart (it looks like an ever-steepening curve; the instructor’s bet is that the recent jump came from reasoning techniques and won’t keep sharpening forever — go check whether that prediction holds), plus coding- and tool-use-specific tests, where the open-source GLM-4.7 from Z.ai scores surprisingly strongly.
You’ve put up with more talking than anywhere else in the course — thank you. You now understand what’s under the hood. Tomorrow: tools, tools, tools — Cursor, Copilot, Codex, Antigravity — and your first real build. You’re ~13% through the journey. 🎉
✓ Key takeaways
- Six workflows run from micromanagement to swarms, across the 2025 and 2026 mindsets.
- Ralph Loops wrap an agent’s loop in an outer feedback loop — run for hours, even overnight.
- Match the workflow to the task: hands-on for mission-critical, let-it-run for greenfield/boilerplate.
- LLMs are a real multiplier, not a flat 10× — and “the LLM wrote it” is never an excuse.
- Bookmark artificialanalysis.ai to compare intelligence, speed and price.