Answers are grounded in this paper and cite their sources.
Large language model (LLM) agents no longer only answer questions; they now call tools, read databases and change the state of business systems. Orchestration frameworks such as LangGraph coordinate these agents well, but they leave execution governance to every application: who may write, whether a decision is still valid at the moment of writing, what happens when a call is retried, and how each effect is traced. This paper surveys the research that addresses parts of this problem: reasoning-and-acting agents, reliability and safety benchmarks, prompt-injection defences and guardrails, privilege control, transaction and compensation models, observability, and AI-risk governance standards. From this survey we derive a taxonomy of seven failure classes that occur on an agent’s write path and show that no single existing approach covers stale decisions, unauthorized actions, duplicate side effects and incomplete audit together. We then present AgentForge, a domain-agnostic tool-effect gateway built on the idea of harness engineering. In AgentForge an agent can only propose an action. The gateway validates it against typed contracts, current business state and declarative policy; obtains human approval bound to a cryptographic state fingerprint; enforces exactly-once effects with idempotency keys and optimistic concurrency; revalidates the action immediately before commit; verifies the post-condition; and records the whole journey as an OpenTelemetry trace. The framework is demonstrated on a simulated supply-chain procurement workflow. We specify a paired evaluation protocol (60 scenario templates × 3 runs per arm, judged on the final database state) with measurable targets: at least 90% fewer unauthorized commits, at most 1% duplicate effects, at least 95% of stale approvals blocked, at most 5% false blocks, at least 98% audit completeness and below 500 ms median overhead. We close with open challenges and a research agenda for trustworthy agentic infrastructure. Index Terms Agentic AI, LLM agents, harness engineering, tool-effect gateway, execution governance, policy engine, human-in-the-loop approval, state fingerprint, idempotency, optimistic concurrency control, audit trace, OpenTelemetry, procurement, LangGraph.
No takes yet. Share an insight, caveat, or question.
Veer et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: