Ninety four percent. That is how much Vivek Haldar cut token usage on a recurring agent workflow, once he stopped asking a language model to replan the same routine every single day.
Haldar, a software engineer who blogs at vivekhaldar.com, described the change in a post published July 14. His workflow scans his blog archive, checks whether a post has run recently, and drafts a LinkedIn update linking back to it. He never lets the draft post automatically.
The original version lived entirely inside a natural language Agent Skill: instructions telling an agent, currently OpenAI’s Codex, where to look, what counted as a good candidate, and what shape the final draft should take. Every run meant the model interpreting those instructions from scratch, planning steps, calling tools, and tracking state.
That approach paid off early. Haldar could describe the workflow in plain language and adjust it without touching code. But after enough repeated runs, the paths it took stopped varying. It searched the same sources, built the same inventory, and applied the same filters every time, work that carried no exploratory value once discovered.
So Haldar rebuilt it. The natural language skill became what he calls a thin bootloader: a short trigger that hands off to a Python program. That program fetches sources, builds the inventory, checks recent posts, and applies filters, all without a model call. The result was 94 percent fewer tokens and 87 percent lower latency, with output quality he judged unchanged across his runs.
Only two steps in the workflow still call a model: picking a good candidate from the filtered list, and writing the draft itself. Those are the parts that depend on reading meaning into text, not on repeating a known procedure.
Haldar built the new version by feeding a capable model his historical execution traces alongside the original skill description. The traces showed what the agent had actually done on past runs, including branches and state the natural language version never made explicit. That let the model identify which steps had become mechanical and which still needed judgment.
The mechanism generalizes past one blogger’s LinkedIn habit. Any recurring agent workflow accumulates traces once it runs enough times. Those traces are what expose which parts are actually stable. Haldar frames the natural language skill as a specification and the traces as the operational detail that gets compiled into code once the shape of the task stops changing.
He also points to a reason model vendors are unlikely to push this pattern themselves. Their revenue scales with token consumption, and a technique that cuts tokens by 94 percent while holding output quality steady works against that incentive directly.
For teams running any agent skill daily or weekly, Haldar’s sequence is concrete. Ship the natural language version first. Gather traces across real runs, then compile whatever has stopped changing into code and keep the model on only the steps that require judgment. Skip the compile step and every run keeps paying frontier model prices to replan a procedure that was already solved weeks earlier.
Vivek Haldar detailed the token reduction on his blog, vivekhaldar.com, in a post published July 14, 2026.
