Mario Zechner's agent workflow is two sentences long
The reply was a joke about prompt scaffolding, not a benchmark, but it landed in a week full of elaborate agent machinery.
Asked how he works with agents, Mario Zechner answered "i tell them what to do. when they do it badly, i yell at them". He then posted the same line again as "how i work with agents in 2026" with the instruction to ignore all other guides on here.
i tell them what to do. when they do it badly, i yell at them.
how i work with agents in 2026. ignore all other guides on here.
That is the whole thing. There is no repo, no measurement, no before and after. It is a jab at the genre of agent workflow guides, delivered in the format of an agent workflow guide, and it should be read as a joke rather than as a result.
Why it lands anyway
The joke works because the opposite position has been very loud. In the same stretch, engineers we follow have been publishing serious machinery around agents. Matt Pocock wired up a pre-commit hook that blocks any agent-authored file over 5,000 tokens. Theo built a council skill that makes several models argue before a merge. dax keeps one long-running pinned session per project and farms work out to subagents. Anthropic, for its part, has started telling companies to delete unnecessary subagents.
So the spread of advice currently on offer runs from a hard token ceiling enforced at commit time to a man saying he yells. Both ends claim to be shipping.
For working engineers
The useful read is not that scaffolding is pointless. Zechner did not test that, and nothing in his two posts offers evidence either way. The useful read is that scaffolding on its own is clearly not the variable that separates people who get work out of agents from people who do not, because you can find productive engineers at both extremes of how much structure they wrap around the model.
What Zechner's line does describe, if you squint past the comedy, is a tight correction loop. Give the agent a task, look at the output, reject it loudly and specifically when it is wrong. That is the part every one of the more elaborate setups is also trying to automate, whether through a hook, a review council or a subagent that checks another subagent's work. The difference is where the judgment lives, in a human reading diffs or in a rule the agent has to pass.
Nobody has a controlled comparison of the two approaches. What exists right now is a pile of individual reports, and one engineer saying his pile is two sentences long.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.