How can engineering teams eliminate runtime deadlocks and operational failures in autonomous agents without gutting workflow utility?

Large language model agents are expanding across software engineering, scientific discovery, customer service, and enterprise automation to manage multi-step tool calls where real-world effects are difficult or impossible to reverse. Yet operating these systems creates an operational friction point between achieving the target outcome and enforcing domain rules. Standard safety interventions prevent unauthorized actions by halting the agent, but this intervention frequently leaves autonomous processes paralyzed midway through execution.

Formal verification of hierarchical plans prior to tool execution resolves the trade-off between safety enforcement and practical utility. As Mike He, Sharad Malik, and Aarti Gupta from Princeton University alongside Ankush Desai from Snowflake presented in research for the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), SOTERIA is a framework that pairs pre-execution formal checks with active runtime enforcement. Published under arXiv identifier arXiv:2611.17880v1, the architecture establishes verified execution paths before invoking external systems.

The Breakdown of Standard Guardrails

Deploying autonomous agents at scale exposes structural deficiencies in standard runtime oversight. Current enterprise workflows generally implement one of two protective mechanisms. The first approach relies on secondary language models acting as evaluators to inspect every proposed action before dispatch. These secondary models are straightforward to integrate, but their evaluations remain probabilistic and fail to provide absolute safety guarantees.

The alternative relies on specification-based monitors. These monitors evaluate preconditions, postconditions, or trajectory invariants directly at runtime, delivering stronger verification boundaries. However, when an invalid call is blocked, the monitor provides no systematic mechanism to recover. As the research paper notes:

"These objectives are in tension – blocking an unsafe action prevents a violation but can leave the agent stranded with no principled way to recover."

Halting an operation leaves the agent in a non-recoverable state, directly suppressing task completion rates across customer-facing or back-office tasks. The industry assumed that enterprises had to accept either unverified probabilistic risk or constant operational downtime.

Can autonomous agents enforce domain policies without sacrificing task completion?

Formal Verification and Runtime Recovery

SOTERIA answers that problem by establishing verified hierarchical paths before invoking any external tool. The framework relies on a one-time setup phase that converts natural-language policy documents and tool definitions into formal policy specifications, tool contracts, and reusable sub-plans describing standard domain procedures. These specifications and tool contracts are generated with LLM assistance and then manually validated against domain policy.

During runtime planning, the system composes hierarchical plans from its library of pre-verified sub-plans and verifies them against the specifications. Plan verification checks three properties: tool contracts to guarantee every call satisfies its precondition, progress to verify the plan cannot get stuck, and policy compliance to ensure executions follow domain rules. When verification checks fail, targeted diagnostics instruct the planner to repair and re-verify the sequence within a fixed retry budget. If full verification cannot be completed, the framework proceeds using the strongest partially verified plan as an operational fallback.

Execution relies on an integrated runtime monitor that inspects each tool invocation against formal specifications and blocks unauthorized actions. Unlike unguided systems that stall upon a rejected action, an executor in this framework consults the verified plan for alternatives and recovers along a guided path.

Quantifying Operational Reliability

Empirical evaluations demonstrate that formal planning fundamentally shifts agent reliability metrics across diverse domains.

At the same time, verified planning eliminated the productivity penalties typical of conservative guardrails. As the assessment shows, the framework delivered 1.7 times the task success rate of the best-performing baselines without recording a single policy violation.

Engineering teams managing enterprise agents face escalating costs when runtime failures stall automated operations or breach compliance policies. Bridging pre-execution contract verification with runtime enforcement proves that policy compliance does not require operational sacrifice. The transition toward verified planning libraries establishes a durable technical path for multi-step enterprise workflows.

Side by side

A one-time manual review of formal specifications and subplan libraries shifts the entire enforcement burden from unpredictable runtime environments straight to the design stage. Traditional runtime monitors freeze faulty workflows because they lack systemic recovery mechanisms to escape a dead end. Unlike secondary evaluators with probabilistic guesses, upfront hierarchical planning builds execution paths around strict tool contracts to guarantee safe progress before external systems ever run. Autonomous enterprise agents still rely on engineers by hand mapping out safety contracts step by step before production.

AI AgentsAutonomous AgentsArtificial IntelligenceAutomationWorkflow Automation