Snap Inc. has released Agent Format (agentformat.org), an open declarative YAML standard designed to let developers define AI agents once and run them on any compliant runtime without code changes. The v1.0 specification, published under Apache 2.0, grew out of a concrete internal problem: three separate Snap engineering teams were independently building agents using LangChain, Google ADK, and custom Python scripts, and could not share tooling, governance policies, or agent definitions across those frameworks. Each runtime migration required rewriting agent code from scratch. The solution was a vendor-neutral .agf.yaml file that decouples an agent's cognitive blueprint — its identity, I/O contract, tool access, constraints, and execution strategy — from any particular runtime implementation.
The schema is grounded in Partially Observable Markov Decision Process (POMDP) formalism, giving it a mathematically rigorous foundation that distinguishes it from the graph-of-nodes models common in most agent platforms. The "tighten-only invariant" for multi-agent systems ensures child agents can never exceed the constraints of their parent, making authorization policies strictly more restrictive as orchestration depth increases. The format also separates agent-level constraints from organization-level policies, with the runtime composing both at execution time — meaning compliance auditors can verify behavior by reading YAML rather than framework-specific code.
The standard ships with a JSON Schema for validation, a CLI linting tool (agf lint), multi-language SDKs for Go, Python, Java, and TypeScript, and an interactive browser playground. A three-tier conformance badge program covering Parse, Execute, and Full Governance capabilities gives runtime vendors like <a href="/news/2026-03-14-nanoclaw-partners-with-docker-for-hypervisor-level-agent-sandboxing">NanoClaw</a> a path to adoption without requiring a central registry.
Agent Format is explicitly positioned as one layer of a three-part production AI stack alongside the A2A Protocol (agent-to-agent communication, led by Google) and the Model Context Protocol (agent-tool use, originated by Anthropic). The boundary between these layers is not entirely settled — A2A's agent card format already encodes some identity and capability metadata that overlaps with Agent Format's scope — but Snap has staked out the definition layer as its distinct concern. The governance structure is modeled directly after the OpenAPI Initiative and CNCF, with a vendor-neutral Standard Committee using lazy consensus for minor changes and a supermajority requirement for breaking changes.
The strategic parallel Snap draws is to Kubernetes: just as Kubernetes manifests decoupled infrastructure intent from imperative shell scripts without killing any particular container runtime, Agent Format aims to become the lingua franca above LangChain, AutoGen, PydanticAI, and Google ADK — sitting on top of them, not replacing them. The Apache 2.0 and CC-BY-4.0 licensing, combined with explicit attribution to "Snap Inc. and The Agent Format Authors," signals intent to transfer community ownership — a deliberate contrast to Google ADK and Microsoft AutoGen, which remain under direct corporate stewardship. The primary risk is adoption velocity: if major cloud providers ship proprietary agent definition formats before the Standard Committee achieves a critical mass of compliant runtimes, the window for a neutral standard could close before portability moves from specification to practice.