Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-24 · high
It rewrites plain-English agent instructions so each step names the files it reads and writes, then measures the gain on 488 real tasks.
Why now: The paper posted to arXiv in August 2026, as agent frameworks keep piecing together workflows from plain-English task lists.
A compiler called Artic takes natural-language agent instructions and rewrites them into steps that name exactly which files or data objects they read and which ones they produce. The paper's case is that plain-English workflows fail as agent instructions because the data dependencies between steps stay unstated, so the executor is left guessing which prior result a given step actually needs. Compiled steps fix that by declaring their inputs and outputs up front, per the Artic paper.
Two more pieces of the design stand out. Constraints gate the artifacts a step produces, catching bad output before it flows into the next step. Control transfers between steps are explicit instead of inferred, which also surfaces steps that depend on more of the workflow's state than they should.
Tested across 488 problem instances in 11 real domains, the compiled version raised the task resolve rate by 28 percentage points over the natural-language baseline. The paper doesn't say which domains saw the biggest gains, or whether the compilation step adds meaningful latency to a workflow run.
Twenty-eight points is a big jump for what's a rewrite pass on the instructions, not a new model or more training. If that number holds up outside this paper's own domains, it argues that agent frameworks lose more reliability to unstated data dependencies than to model capability, and that a compiler sitting between the prompt and the executor is worth more than a bigger context window.
Each link below shares sources, entities, or timing with this story.
Shared entity: Measured / Same source domain / Shared topic / Earlier coverage
Both cover Measured; reported by the same outlet (arxiv.org); overlapping topics (agent, step, workflow).
Shared entity: Compiled / Shared topic / Earlier coverage
Both cover Compiled; overlapping topics (agent, artifact, step); earlier Compiled coverage from 2026-07-28.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, constraint, data); pushes against this story (versus).
Reported by the same outlet (arxiv.org); overlapping topics (agent, data, declare); pushes against this story (but).
Shared entity: Measured / Same source domain / Earlier coverage
Both cover Measured; reported by the same outlet (arxiv.org); earlier Measured coverage from 2026-08-16.
Both cover Measured; reported by the same outlet (arxiv.org); earlier Measured coverage from 2026-08-14.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (agent, workflow); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (agent, artifact); pushes against this story (versus).