source: techcrunch ai: microsoft offers devs a better way to control ai agent behavior
level: technical
microsoft has released an open-source standard called agent control specification, or acs, to help developers manage what ai agents can and cannot do. the spec lets teams write policy files that define allowed actions, forbidden actions, when human approval is needed, and what evidence to log. these policies are checked at several points during an agent's workflow, such as before receiving input, before calling a tool, after a tool returns a result, and before sending a final response.
currently, developers use a mix of system prompts, custom code checks, and classifiers to control agents. these methods work but often lead to fragmented controls that are hard to audit and reuse across different frameworks. acs aims to unify these controls into a common governance layer. policies can allow, block, redact, or escalate actions for human approval. developers can also add classifiers and language models as judges to evaluate inputs and outputs.
the specification ships as a software development kit with plugins for popular frameworks like langchain, openai agents sdk, anthropic agents sdk, autogen, crewai, and semantic kernel. because policies are single files, they can be bundled with agents, letting security rules follow an agent across different systems. this helps enterprises deploy agents more safely by ensuring consistent behavior and easier auditing.
why it matters: it gives ai and data science teams a reusable, auditable way to enforce safety and compliance rules across agent workflows, reducing the risk of unintended actions.
source: techcrunch ai: microsoft offers devs a better way to control ai agent behavior