ContrAgent Uses Logic Contracts to Gate and Audit LLM Tool Calls
A 16 September 2026 arXiv paper argues assume guarantee contracts compiled to automata can gate agent actions online and audit traces offline with deterministic verdicts and lower latency than LLM judges.
What changed
Researchers posted ContrAgent: Symbolic Temporal Supervision of LLM Agents Using Contracts to arXiv on 16 September 2026 (arXiv:2609.18128). The framework captures an agent's behavior as tool call traces, formalizes required behaviors with assume guarantee contracts in linear temporal logic over finite traces (LTLf), and compiles each contract to a deterministic finite automaton. That automaton gates actions online during execution and evaluates recorded traces offline for audit.
The contract library acts as a reusable knowledge base independent of the underlying model, applicable across agents in the same task domain. On four benchmarks spanning both online gating and offline evaluation, ContrAgent matches state of the art LLM judge and rule based guardrail baselines while producing deterministic reproducible verdicts. In online mode the authors report orders of magnitude lower per call latency than stochastic LLM judges.
Why it matters
Enterprise agent rollouts are stuck between slow LLM judges that vary run to run and brittle regex guardrails that miss multi step violations. ContrAgent offers a middle path: policies written as temporal contracts compile to deterministic gates that security and compliance teams can review before deployment. The same artifact supervises live tool calls and post hoc audits, which reduces the usual split between runtime middleware and forensics pipelines.
For regulated workflows (finance approvals, healthcare data handling, privileged file access), deterministic verdicts matter as much as accuracy. Matching judge quality at lower latency is the operational unlock for high throughput agent fleets.
Who is affected
Agent platform architects, compliance engineers translating policy into enforceable rules, security teams tired of non reproducible LLM judge evaluations, and procurement leads comparing guardrail vendors on auditability not marketing claims alone.
What to do next
Identify one high risk agent workflow with explicit sequential rules (for example: fetch customer record, then mask fields, then call summarizer) and prototype LTLf contracts that an automaton can enforce before expanding to library scale governance.
What to watch
Open source release of ContrAgent contract libraries for common enterprise domains, and head to head evaluations against trajectory guards such as HazardAuditor on multi turn tool use attacks where deterministic contracts may miss semantic hazards.
Sources
- Primary. arXiv, Symbolic Temporal Supervision of LLM Agents Using Contracts (16 September 2026). ContrAgent design, LTLf compilation, benchmark parity, and latency claims.