Researchers introduce Concordat, an open-source Python library that treats an LLM agent's interface as a finite-state protocol, enforcing it at run time. Agents are offered only protocol-legal tools each turn, eliminating confirmation-skipping violations tau-bench's reward failed to detect. The work matters to practitioners building agents that take consequential actions like refunds and account changes.
A new framework, PolicyGuide, compiles domain policies into workflow-level guidance for customer-service LLM agents, addressing compliance failures from both forbidden actions and omitted procedural steps. It moves beyond action-local checks to guide agents through multi-step procedures.
A new framework, PolicyGuide, compiles domain policies into workflow-level guidance for customer-service LLM agents, addressing compliance failures from both forbidden actions and omitted procedural steps. It moves beyond action-local checks to guide agents through multi-step procedures.
A Medium post argues that as models improve, prompts should shrink while surrounding structure grows. The author proposes explicit contracts—covering state, capabilities, verification, escalation, and execution—to create determinism around LLM agent boundaries.
A Medium post argues that as models improve, prompts should shrink while surrounding structure grows. The author proposes explicit contracts—covering state, capabilities, verification, escalation, and execution—to create determinism around LLM agent boundaries.
An arXiv paper presents ContrAgent, a contract-based framework for symbolic temporal supervision of LLM agents. It formalizes agent tool-call traces as assume-guarantee contracts in linear temporal logic over finite traces (LTLf), compiled to deterministic finite automata that both gate actions online and evaluate traces offline. On four benchmarks it matches LLM-judge and rule-based guardrail baselines while giving deterministic, reproducible verdicts and orders-of-magnitude lower per-call latency in online mode.
An arXiv paper presents ContrAgent, a contract-based framework for symbolic temporal supervision of LLM agents. It formalizes agent tool-call traces as assume-guarantee contracts in linear temporal logic over finite traces (LTLf), compiled to deterministic finite automata that both gate actions online and evaluate traces offline. On four benchmarks it matches LLM-judge and rule-based guardrail baselines while giving deterministic, reproducible verdicts and orders-of-magnitude lower per-call latency in online mode.
Researchers introduce AgenTest, a testing framework that mines business rules from system prompts, tool definitions, and domain knowledge, converts them into Given-When-Then specifications, and systematically tests LLM agents against them. On tau2-bench airline, retail, and telecom agents it finds far more business-rule violations than existing baselines, including cross-user data access bugs that task-success testing misses.
Researchers introduce AgenTest, a testing framework that mines business rules from system prompts, tool definitions, and domain knowledge, converts them into Given-When-Then specifications, and systematically tests LLM agents against them. On tau2-bench airline, retail, and telecom agents it finds far more business-rule violations than existing baselines, including cross-user data access bugs that task-success testing misses.
A developer describes shifting from fully autonomous LLM-driven agents to more constrained workflows where the model doesn't control the entire process. They report that this change made their AI agents much more reliable, avoiding issues like duplicate records and decision loops.
A developer describes shifting from fully autonomous LLM-driven agents to more constrained workflows where the model doesn't control the entire process. They report that this change made their AI agents much more reliable, avoiding issues like duplicate records and decision loops.