← News

Agents & MCPSources: arXiv · 19 Nov 2026

Intent contracts separate planning from tool execution in LLM agents

A new paper proposes an authorization design where the agent's planner never touches tools directly. A small trusted model writes each tool call after code validates a short logged contract, keeping the planner's influence narrow.

In plain words

Many AI agents let one large model both decide what to do and run the tools. This paper splits those jobs: the planner only files a short record stating the goal and quoting the user, then a separate trusted model writes the actual tool call. In tests this made agent actions easier to trace, though it did not stop all mistakes.

What was announced

The authors describe intent contracts as an authorization interface for tool-using agents. The planning model holds no tools; each step it files a logged contract containing a goal, a verbatim quotation of the user, and scope fields. Validation code checks the contract, and a small trusted model compares it with the user's words and writes the tool call itself. If that validation and dispatch code is correct, every executed call comes from the trusted model, and the planner's influence is limited to the contract logged beforehand.

In benchmark tests using gpt-6.1-terra as planner and GLM-5.3-Flash as the trusted model across three AmPermBench task families, every object changed under contracts was named in the corresponding contract's goal, while planner-written commands under review rarely named objects in goals. When the planner added unrequested work, contracts finished the user's task in 27 of 30 runs versus 16 of 30 for an intent-visible reviewer, which can only reject an entire command. The authors caution that contracts did not reduce unauthorized actions: once environment terms were clarified, conditions that changed state hit a wrong target in 40-49% of runs, an error tied to the planner's handling of ambiguous references. The over-reach evaluation used one model pair and no adversarial planners.

The paper is represented here by its abstract, so full methods and results are not covered. The authors state they are releasing the plugin and run records.

What it does not show

  • Evidence is limited to the paper's abstract; full experimental detail is not available here.
  • Benign over-reach was tested with a single pair of models and no adversarial planners, per the authors.
  • Contracts did not reduce unauthorized actions, and wrong-target errors of 40-49% persisted in state-changing runs.

For security and IT teams

Scope
The design assumes the validation and dispatch code itself is correct; the security guarantees described depend on that assumption.

Sources

Source-support review completed 20 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 20 Nov 2026

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