Skip to content

Convergence Thesis Discussion Draft

Use this as the draft for the pinned GitHub Discussion.

Suggested Title

Why coding agents need an operational layer, not just better prompts

Suggested Category

General

Draft Body

The last few months made one thing clearer for me:

coding agents do not fail because they lack raw generation power.

They fail because the system around them is weak.

They forget what was learned last week. They repeat failed approaches. They do not challenge their own plans. They do not naturally turn finished work into better next work.

That is the gap AgentOps is built to close.

Publicly, the framing is simple:

AgentOps is the operational layer for coding agents.

It gives them:

  • Bookkeeping so important learnings, findings, and reusable context do not disappear between sessions
  • Validation so plans and code get challenged before they ship
  • Primitives so operators can pull one capability at a time
  • Flows so those capabilities compose into repeatable work

The outcome is the part I care about most:

work compounds instead of resetting every session

That framing did not come out of nowhere. It converged from three directions:

  1. Karpathy's LLM wiki framing The tedious part is not reading. The tedious part is the bookkeeping.
  2. Block / Owen Jennings The moat is the feedback loop around the markdown file, not just the model.
  3. What we learned building AgentOps in production Agents are not the product. The system around them is.

That is why I no longer think the right product category is "workflow tooling" or "memory for agents."

Those are pieces.

The real category is the operational layer:

  • bookkeeping
  • validation
  • composable primitives
  • linked flows
  • a repo-native feedback loop that makes the next session better than this one

Technically, the best internal framing is still:

AgentOps is a context compiler.

Raw session signal becomes reusable knowledge, compiled prevention, and better next work.

But that is the technical reveal, not the public headline.

The public headline is simpler:

coding agents don't do their own bookkeeping

If you are building with coding agents every day, that is the pain I care about.

If this resonates, I’d especially like feedback on three questions:

  1. Where do your coding agents still reset from zero?
  2. What bookkeeping do you still do manually?
  3. Which part is most missing in your stack today: bookkeeping, validation, primitives, or flows?

Publish Checklist

  • tighten wording for your current voice before posting
  • add one repo screenshot or nightly dream-cycle artifact link after na-gtm.7 is live
  • pin after publish