Build or buy an agent mesh

Updated 19 November 2026 · Editorial assessment

What does building an agent mesh from open parts involve?

Building means you assemble protocol SDKs, a proxy and an agent runtime, and then you run them. The protocol layer is the most settled part. The A2A project offers official SDKs for Python, Go, Java, JavaScript, C#/.NET and Rust, and every SDK implements the same A2A v1.0 specification.

The Python SDK implements specification 1.0, with a compatibility mode for 0.3, and installs with pip install a2a-sdk or uv add a2a-sdk. The Java SDK is a multi-module Maven library with client and server support over JSON-RPC, gRPC and REST transports.

For the proxy, agentgateway is an open source data plane that provides connectivity for agents, MCP tools, LLMs and inferences. Its README describes an A2A gateway to enable secure agent-to-agent communication with capability discovery, modality negotiation and task collaboration. The same README lists JWT, API keys and OAuth for auth, along with RBAC through a CEL policy engine, rate limiting, TLS and OpenTelemetry.

For the runtime, kagent defines an Agent as a Kubernetes custom resource that brings together behavior and runtime configuration. Its controller exposes gRPC, A2A, and MCP APIs to clients, and agents are defined in YAML for GitOps workflows. The kagent site states that kagent and Agent Substrate are Apache 2.0 licensed CNCF Sandbox projects.

Is Solace Agent Mesh an open-source building block?

Solace's Agent Mesh README says Agent Mesh Desktop is for building and testing on your own computer. For production, Solace Agent Mesh requires a Solace commercial license, its README says. The product hosts agents and runs their tools, and it coordinates them over an event broker with sessions, authentication, evaluators and observability built in.

The repository LICENSE is the Business Source License 1.1. It grants the right to copy, modify, redistribute and make non-production use, and it names Apache License 2.0 after four years as the change license. The README states that production use requires a commercial license from Solace. It also says the source code is coming soon, and that until then the repository holds Agent Mesh Desktop downloads and examples.

The product page says it is always free for smaller companies, meaning individuals and companies making less than $25M/year in revenue. Ask Solace how that free use and the commercial license term apply to your company before you plan around either. The Solace Agent Mesh profile covers the layer it occupies.

What does the A2A protocol leave for you to build?

A2A fixes how agents describe themselves and exchange messages. Discovery, identity sources, authorization and API management stay with your team, and each one becomes a requirement in a build.

What does a bought layer add?

A bought layer moves some of the operating work into a vendor contract, but it does not remove all infrastructure. Solace Agent Mesh is one example. It describes itself as an agent development and runtime platform that supports build, test, deploy, monitor and improve. For production you can use Agent Mesh Cloud, a managed cloud service, or deploy it to Kubernetes.

For architects serving several business units, the Solace docs describe namespace separation. Each deployment subscribes only inside its own namespace, so teams can share one event broker without seeing each other's traffic. The same page says a shared or production deployment must run against a Solace event broker.

Gateways, registries and hyperscaler runtimes are separate layers that sit beside a mesh. If you already run one of them, buy only for the gap it leaves.

Which costs grow with agent count?

Registry entries, client configuration, telemetry and idle compute all grow as you add agents.

What running state does a build have to keep?

Plan for databases from day one. In kagent, PostgreSQL stores sessions and tasks along with conversation history, so they persist apart from the running agents.

In the A2A JavaScript SDK, the in-memory stores lose all state on restart. Its database stores support PostgreSQL, MySQL, SQLite and Cloudflare D1, and the README warns that push notification credentials are stored in plain text.

Budget for upgrades as well. The roadmap section of the kagent README says kagent is currently in active development.

What questions should decide build or buy?

Work through these in order and record the answers in your RFP checklist.

  1. Name the team that owns the registry and fixes a wrong Agent Card.
  2. Decide where identity comes from for calls between business units and clouds.
  3. Estimate the agent count in two years and list which of the costs above grow with it.
  4. Have legal confirm the production license terms, including any Business Source License terms.
  5. Check whether a gateway or runtime you already own covers part of the need, so you buy only the missing layer.
  6. Assign on-call ownership for upgrades of the SDKs, the proxy, the database and the broker.

What do enterprise buyers ask?

Do I need an agent mesh if I only use A2A SDKs?

Direct configuration is used for tightly coupled systems, private agents and development work, and the A2A docs call it straightforward for known, static relationships. Once agents and teams change often, the registry and identity work described above becomes the main project.

Must A2A traffic use HTTPS?

Yes, in production. The documentation states that production A2A communication must use HTTPS.

Can kagent run agents built elsewhere?

Yes, if the agent fits kagent's contract. A Harness can run any container image that speaks kagent's A2A contract, as well as kagent's own engine, Claude or Codex.

Who governs the A2A protocol?

A2A is a Linux Foundation project, contributed by Google. The A2A protocol guide covers the specification in more depth.

Which sources does this page cite?

Reviewed Nov 2026