Intent contracts keep tool calls out of the planning model
Original title: Intent Contracts: Taking Tools Away from Agents
On AmPermBench, contracts finished the user's task in 27 of 30 over-reach runs, versus 16 of 30 for an intent-visible reviewer.
The paper proposes intent contracts for tool-using LLM agents: the planner holds no tools and submits a logged goal, a verbatim user quotation, and scope fields. Validation code and a small trusted model check the contract and write the tool call. With gpt-6.1-terra as planner and GLM-5.3-Flash as the trusted model, every object changed under contracts was named in the producing contract's goal. Contracts did not reduce unauthorized actions and remained indistinguishable from blind review; after clarifying environment terms, 40-49% of state-changing runs hit a wrong target. The authors release the plugin and run records.
Unconfirmed: The study evaluates benign over-reach with one model pair, not adversarial planners.
- Contracts completed the user's task in 27/30 runs when the planner added unrequested work.
- The intent-visible reviewer completed that task in 16/30 runs because it can only refuse a whole command.
- Wrong-target errors occurred in 40-49% of state-changing runs after environment terms were clarified.
- Evaluation used one model pair and benign over-reach, not adversarial planners.
- Products
- AmPermBench
- AI models
- gpt-6.1-terraGLM-5.3-Flash
From the original article
1 min read at arxiv.orgIntent contracts are an authorization interface for tool-using LLM agents in which the planning model holds no tools. For each step it submits a short logged contract: a goal, a verbatim quotation of the user, and scope fields. Code validates the contract, and a small trusted model checks it against the user's words and writes the tool call itself.
By Ziyuan Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.