Review the short alignment doc, not the thousand-line plan
The team that popularized research, plan, implement published what they got wrong. Read the code, and put human attention on the short doc before the plan, where a bad assumption is still cheap to fix.
Written by
-
Benedikt Stemmildt Founder & Co-CEO
Published: 21 May 2026 Updated: 2 min read
The team that popularized “research, plan, implement” for coding agents just gave a talk on what they got wrong about it. Two corrections, both about one thing: where the human looks.
The early advice did not hold
The early advice was: don’t read the code, read the plan. The plan is the spec, the code is just the output, let the agent run.
They worked that way for six months. It didn’t hold. They had to rip out and replace large parts of what got built.
So here’s the corrected version. Read the code. And the part I didn’t expect: don’t sit and read the long plan either.
Reviewing the plan is the same job twice
The reason is leverage. A thousand-line plan becomes about a thousand lines of code, within ten percent. Reviewing the plan is the same work as reviewing the code, and the code always drifts from the plan a little, so you end up reading both. That isn’t leverage. It’s the same job twice.
The leverage sits one step upstream
Real leverage sits upstream of the plan. Before the plan there’s a short doc, a couple hundred lines: current state, where we’re going, the patterns to follow, the decisions still open. The agent writes out what it found and what it intends to do, and you correct it right there, before a line of code exists.
That’s the cheap place to be wrong. A bad assumption caught in that doc costs two hundred lines to fix. The same assumption caught after the code is written costs a thousand.
It’s still a small factory inside one ticket. Research, then alignment, then plan, then code, each gate in order. What moved is where the human stands: not on the plan, reading the longest document in the process, but one gate upstream, while the thing is still short enough to change your mind on.
And when the code lands, you read the code. 2026 is a bad year to ship work nobody read.
Where does your team’s review actually happen, and is that the cheapest place it could?
Source: Dexter Horthy, “Everything We Got Wrong About Research-Plan-Implement.”
Further reading
Written by
-
Benedikt Stemmildt · Founder & Co-CEO
More on Specification and review
- When code is cheap, the design decision is the expensive part When a full feature costs a few dollars to generate, throwing code away stops being waste and becomes the cheapest way to test an idea. The constraint moves from writing the code to deciding what survives.
- The costly mistake is hovering over your agent Writing 20 lines by hand is no longer the overcorrection to worry about. Staying in the loop so tight that the agent never gets room to work costs far more.
- AI velocity without engineering discipline fires slop A framework sorts tech roles into four archetypes, and it reads like a raid party. AI rollouts without engineering discipline add damage output with no tank and no healer. Agentic engineering balances the party.
Test the constraint in your system
The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.
Training on this topic: Agentic Engineering for Developers and QAWant to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check