Is Agent Executor tied to one agent framework?

No. Google describes AX as harness-agnostic and says it is not an agentic framework.

Does AX support MCP tools and Agent Skills?

The current repository says built-in harnesses such as Antigravity can discover and call configured MCP tools and include Agent Skills support. This is a harness capability, not a universal guarantee for every integration.

Does A2A or MCP compatibility prove that an integration is safe?

No. Protocol compatibility does not prove identity propagation, authorization, data handling, semantic correctness, retry safety, or support for the exact version and feature used.

AI Agent - Aug 12, 2026

Agent Executor Compatibility: A2A, MCP, Skills, Harnesses

Quick answer

AX is designed as a harness-agnostic runtime, not an agent framework. Google’s launch announcement describes interoperability with custom agents, ADK, LangChain and LangGraph, A2A-based agents, MCPs, skills, Google harnesses, and managed or self-built agents. The current repository documents a narrower implementation surface: built-in Antigravity harnesses, a custom HarnessService, configured MCP discovery and calls, and Agent Skills support in built-in harnesses.

Treat those layers carefully. An announcement can establish product direction; the pinned repository establishes what you can inspect and test today. Neither proves that two integrations have equivalent semantics or security.

Compatibility has several meanings

Layer	Passing test	It still does not prove
Build compatibility	Packages and generated code compile	Runtime behavior or stable API
Protocol compatibility	Messages can be exchanged	Shared meaning, error handling, or authorization
Harness compatibility	A harness implements the expected service	Correct tool use, recovery, or state lifecycle
MCP compatibility	Tools are discovered and callable	Trust, permissions, schema safety, or idempotency
Skill compatibility	Instructions and assets load	Provenance, prompt safety, or business correctness
Model compatibility	Credentials and requests work	Capability parity, determinism, cost, or data policy
Compute compatibility	An actor starts and resumes	Isolation, scale, support, or migration safety

Use “compatible” only with the layer, version, and test named.

Custom harness boundary

The repository says a custom harness can implement HarnessService. Its current protobuf defines a bidirectional connection with a start request, optional cancellation, output frames, and one terminal response. Traditional tool-use or workflow agents, and even model wrappers, can be represented as harnesses.

That adapter owns important semantics: how inputs become steps, how tool calls and results are represented, how cancellation propagates, which state is persisted, and what constitutes completion or failure. Version the adapter beside the AX protocol and test interruption between every transition.

A2A boundary

Google’s announcement says AX supports agents using the Agent2Agent Protocol. Confirm the current repository path, protocol version, transport, authentication, task-state mapping, artifacts, cancellation, streaming, and error conversion before claiming a working integration. Do not infer that every A2A feature is implemented because the runtime is designed to federate agents.

Most importantly, an A2A peer’s identity is not automatically the end user’s authority. Propagate verified identity and reauthorize every action at the receiving agent and system of record.

MCP and Agent Skills boundary

The current README scopes MCP and skills to built-in harnesses such as Antigravity. Configuration makes an MCP tool discoverable; it does not make it trusted. Review server identity, tool schemas, input validation, scopes, destination resources, output handling, timeouts, retries, and prompt-injection paths.

Skills are instructions and resources, not signed policy. Pin their source, review scripts and dependencies, limit execution permissions, and record the exact version used for each accepted run. Treat changes to a skill as behavior changes that require regression tests.

Compatibility acceptance suite

Pin AX, protocol definitions, harness adapter, A2A or MCP version, skill revision, model, and compute layer.
Run a known-answer task with no tools, then one read-only tool.
Exercise malformed frames, missing fields, oversized content, unknown tool names, and invalid arguments.
Interrupt before and after tool dispatch; verify cancellation and resume.
Attempt identity spoofing, cross-tenant identifiers, permission escalation, and malicious tool output.
Swap one component version at a time and rerun the same fixtures.
Record semantic differences, not merely successful transport.
Keep a fallback harness or direct execution path until the integration is stable.

For runtime lifecycle, read the AX Preview risk guide. For tool and tenant controls, continue with AX security and isolation. Use the readiness checklist for the final gate.

Frequently asked questions

Is AX tied to one framework?

No. It is designed as a harness-agnostic runtime and explicitly says it is not an agent framework.

Does AX support MCP and skills?

The current repository documents MCP and Agent Skills support in built-in harnesses such as Antigravity. Verify the exact harness and revision.

Does protocol compatibility prove safety?

No. Identity, authorization, semantics, data policy, retries, and version support remain separate tests.

Official sources

Source check: August 12, 2026; AX revision 2bcc1637b3c106c16963c5c5464aedb46c6da031. Recheck integrations, supported harnesses, protocol versions, roadmap, and examples before use.