AI Coding Weekly

dax put every subagent on DeepSeek Flash, Opus stays the boss

The model split is cheap. The scoping and proof-of-work rules are the part worth copying.

dax says his team spent the week switching all their subagents to DeepSeek Flash and running Opus 5.5 only on the primary agent. He has tried the arrangement before and it did not go well. This time, in his words, it is working great.

The interesting part is not the price. It is the five rules he listed alongside the setup, which read like the operating manual for anyone delegating to a cheaper model.

What dax actually does

His team runs long sessions per project rather than one-off agent invocations. Opus is told to do all work in DeepSeek subagents, not to pick up the editing itself. When it creates a task it is instructed to be strict about scope. Then two checks. The subagent is asked for proof of work instead of a summary, and Opus does a one step review of the diff to confirm the result is not totally off.

dax's verdict on the orchestrator is short. "opus is amazing at this role", he wrote.

dax
@thdxr
X
ask for proof of work instead of a summary
Nov 6, 2026 · View on X
dax
@thdxr
X
opus is amazing at this role
Nov 6, 2026 · View on X

Why the rules carry the weight

Each of those five lines maps to a specific way cheap subagents fail. Loose scope means a model with less judgment invents work. A summary is the easiest thing for a weak model to produce convincingly, so asking for proof of work, actual output rather than a report about output, removes the cheapest path to looking finished. The one step diff review is the backstop for everything the first two miss, and it is one step on purpose, because a full review by the expensive model would undo the savings.

For working engineers, this is one team's report on their own projects over a week, not a benchmark, and dax is explicit that the same structure did not work for him in the past. What changed between the two attempts is not spelled out. It could be the rules, it could be the models underneath them, and nobody has separated the two yet.

The split-the-tiers pattern keeps showing up. Malte Ubl ran 24 DeepSeek sub agents under an Opus 5.5 orchestrator for about nine dollars in a similar setup. What dax adds is the supervision contract that makes the cheap tier safe to trust, and it costs nothing to copy into your own prompts regardless of which models you are paying for.

Have the good opinion, not the loud one

One email a day. Under three minutes. What working engineers found out about coding with AI.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.