Technical writing in the agentic era
Coding agents can rewrite paragraphs and rebuild interactive charts quickly. When an agent can produce several ways to explain a result, writing more is easy. The scarce work is editorial: deciding what deserves attention, how to show it, and what evidence the reader needs to trust it. The first decision is the order in which the reader encounters that evidence.
Lead with the strongest result
Reader attention is scarce. A long setup can lose people before they reach the main payoff, so put the main result up front. For example, our post leads with this summary:
"We gave an agent a budget of 11,000 H100 hours to find the best data mixture for training a 2B VLM."
Its best data mixture raised the aggregate score from 58.8% to 62.8%, a 6.8% relative improvement over our human-built baseline. It also improved all ten benchmarks. That result belongs near the top. Once readers know what the system achieved, they have a reason to care about its architecture.
Write around visuals
When I skim a technical post or paper, I look at the plots, tables, and diagrams first. They often decide which sections I read. That is why I focus on what the figures show and how readers interact with them.
An accurate plot can still frame a result badly. If progress came in isolated jumps, an ordered experiment plot can communicate the sequence better than a mostly flat progress curve.
Agents make it easy to try another chart, layout, or interactive explanation. We have fewer excuses for making the reader work. PDFs freeze results into static snapshots. Web articles can reveal benchmark descriptions and examples on hover, switch comparisons, or keep explanations beside each figure. The Research Article Template shows this on Hugging Face Spaces with interactive diagrams, math, citations, and PDF export. Inspired by Distill's framework, it treats technical writing as native to the web.
Inspect the output
Add references to earlier work and external announcements, and name and cite every baseline. A technical blog does not need the citation density of a paper. Every figure should still be introduced in the prose rather than left to explain itself.
Watch for passages that move from an eight-GPU node to an unexplained term such as GPU slot. Fluent prose can hide undefined terms or broken reasoning. Apply the same scrutiny to agent-written prose that you apply to agent-written code: trace the logic, define every term, reconcile the numbers, and verify the claims.
Smaller edits matter too: shorten long section headings, give a comparison table a sentence of context, reference every figure from the main text, and connect the end of one section to the start of the next.
My workflow
I start with whichever result has the largest impact or strongest visual and move it near the top. For each other result, I ask the agent for visual options and a recommendation, choose one, and iterate. If the format feels wrong, I switch.
Once the visuals establish the sequence, I write around them. Diagrams and infographics can carry dry background or architecture sections. Then I make a separate pass for citations, factual consistency, transitions, and generated-sounding prose. Finally, I skim without reading every paragraph to see whether the visuals make their point on their own.
The tools are better now. Technical writing should be better too.
