← News

Agents & MCPSources: arXiv · 16 Nov 2026

Co-Sight Harness moves agent workflow control outside the model

A research team reports that long business automation chains can run reliably with the underlying language model left untouched, by enforcing versioned external rules and checks.

In plain words

When automated AI agents handle long business tasks, small mistakes pile up and instructions get misread. The Co-Sight Harness paper describes keeping the AI model itself unchanged while storing the task rules, progress tracking and specialist knowledge as separate, versioned files. A state machine plus machine-checkable gates enforce those files during runs, and people must approve any updates.

What was announced

The authors say two failure modes plague long workflows: errors compound as chains of steps grow, and the vocabulary a step needs differs from the user's phrasing. Their answer is to externalize workflow definitions, state and domain knowledge into versioned control assets, then enforce them at runtime with a state machine and machine-checkable acceptance gates. An independent auditor grades each execution trace, and asset changes require a regression check plus human approval.

On one deployed workflow, a 120-case evaluation reportedly raised unattended completion from 50% to 100%, retrieval recall from 56% to 91.67%, and cut end-to-end latency from 440 to 159 minutes. Two other domains were adopted by teams that did not author the assets, within four and seven days. The paper notes a controlled ablation separating the two failure sources, and the code is published on GitHub.

What it does not show

  • Results come from the paper's abstract only; the full article was not reviewed, so evaluation details cannot be verified.
  • Reported gains cover one domain's 120-case set in detail; the other two domains are described only by adoption time, not performance figures.
  • No independent replication or third-party confirmation of the measurements is available.

For security and IT teams

Scope
The method targets long-horizon business workflows and explicitly leaves the backbone language model unchanged, with code released in the Co-Sight repository.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.