Long-doc AI gap is about reading, not code; subagents close most of it — Contested · resolves 2027-03-31

The question is whether, in work like legal due diligence that means reading thousands of documents, a purpose-built harness (the code around a model that decides how it reads and splits up work) is far stronger than a general AI coding agent.

It decides where vertical AI companies have a moat, because if the edge is only about reading everything, general tools can catch up by changing settings.

The legal AI company Harvey’s own test shows its harness beating general agents by a wide margin; SecondSource reads the gap as coming mainly from general agents reading too little on default settings.

SecondSource’s reading: in long-document professional work, a purpose-built harness beats general agents (AI that plans its own steps and calls tools) mainly on coverage, meaning whether it reads all the material, not on proprietary code.

General agents on default settings read too little, and the gap narrows sharply once subagent dispatch (the lead agent splits the review into pieces for several helper agents) is turned on.

With the right harness, the gap between root models is still about as large as the harness effect itself, and lead agents built by post-training open models have not yet caught up with frontier models.

In Harvey’s self-reported test, the same seven models rose from 23.3% to 62.4% average pass rate with a dispatching harness; its post-trained open lead agent scored 63.0%, below Opus 5 at 77.6%. These figures come from one company’s own benchmark and have not been independently reproduced.

Ahead of AI(Sebastian Raschka)
“to my surprise, Qwen3.6 does actually perform…”

At the due date, check: (1) on comparable long-document due-diligence tasks, whether Claude Code or Codex with subagent dispatch (or dynamic workflows) switched on still trails a purpose-built recursive harness by 20 points or more; (2) whether a post-trained open-model lead agent (such as Harvey's GLM-5.3) matches or beats the best frontier model (Opus 5 at 77.6%) on a held-out set under the same harness; (3) whether a third party redoing the test with real data rooms or lawyer grading finds that the link between coverage and score disappears.

What would overturn it

If (1) holds, the claim that the gap is mainly about not reading everything is withdrawn and the claim that general agents are the wrong default is restored; if (2) holds, the claim that post-trained open models lose to frontier models is withdrawn and the second half of the original anchor is restored; if (3) holds, the coverage mechanism is withdrawn.