Do you need an agent mesh yet? A self-check calculator
Updated 18 November 2026 · Editorial assessment
What does this calculator work out?
It returns two numbers. The first is the number of pairwise links if every agent may reach every other. The second is a readiness level from 1 to 3. Level 1 suggests shared infrastructure can wait. At level 2 it is time to write requirements, and level 3 treats shared agent-to-agent infrastructure as a current need.
The levels are this guide's rules of thumb and carry no standing as a standard. Use them to decide how soon to start an evaluation, then continue with the RFP checklist.
What do you enter?
- Agents in production (agents) Count agents that run against real work, and leave out prototypes. Default 12, allowed range 0 to 1000.
- Frameworks or vendors (count) Count distinct frameworks, runtimes or vendor platforms your agents are built on. Default 3, allowed range 1 to 50.
- Owning teams or business units (count) Count the groups that each answer for at least one agent. Default 4, allowed range 1 to 100.
- Agents cross a company boundary (0 or 1) Enter 1 if any agent exchanges work with an agent run by another company, otherwise 0. Default 0, allowed values 0 or 1.
- Regulator or auditor asks who authorised agent actions (0 or 1) Enter 1 if a regulator or auditor has asked this, otherwise 0. Default 1, allowed values 0 or 1.
How is the result calculated?
The link count is the number of distinct pairs among your agents. Each pair is a connection someone may have to build and secure. The A2A project says that without a shared protocol, each interaction requires custom, point-to-point solutions with significant engineering overhead, and that systems become difficult to scale as agents and interactions grow.
The readiness level uses these rules, applied in order:
- Level 3 applies when you run at least two frameworks and have three or more teams, or when any agent crosses a company boundary.
- If level 3 does not apply, the result is level 2 when any one of these holds: five or more agents, more than one framework, more than one team, or an auditor asking the authorisation question.
- Every other case is level 1.
links = agents * (agents - 1) / 2
level = 3 if (frameworks >= 2 and teams >= 3) or crossCompany = 1
level = 2 else if agents >= 5 or frameworks >= 2 or teams >= 2 or auditQuestion = 1
level = 1 otherwiseWorked example
With the defaults of 12 agents, 3 frameworks, 4 teams, no company boundary and an auditor asking, the link count is 12 x 11 / 2 = 66. The level is 3, because 3 frameworks and 4 teams meet the first rule even though no agent crosses a company boundary.
What does the result not tell you?
The arithmetic is illustrative and uses only the numbers you enter. In most estates many pairs of agents never exchange work, so treat 66 as the ceiling for what you might build.
The levels do not measure risk. A single regulated workflow with two agents can matter more than a large internal estate, so adjust the inputs to your own case.
The model does not name a product. Gateways, registries, event meshes and cloud agent runtimes cover different layers, and the category map shows how they fit together.
- Pairwise links (links)
- 66
- Readiness level (level 1 to 3)
- 3
What do enterprise buyers ask?
Does level 1 mean I should buy nothing?
Usually it means shared infrastructure can wait. BAND (band.ai) gives similar advice in its documentation, which says that for a single agent with no handoff or review, or for a fixed workflow inside one framework, you should keep the simpler setup.
What should I look at first at level 3?
Start with who owns each agent and how agents are identified, using the agent identity guide. BAND's documentation says it fits work moving between agents that do not share one runtime, machine, or owner. Its pages also state limits. BAND routes mentions but does not decide them, so if an agent's prompt or code never tells it to mention another participant, no handoff happens. After a process restart, rehydration limits and provider-session recovery vary by adapter.
Is an agent mesh the same as orchestration?
They overlap but serve different needs. Orchestration usually sequences work within one design and one owner, while a mesh connects agents that different teams or vendors built separately. The page on agent mesh versus orchestration covers the difference.
Which sources does this page cite?
Reviewed Nov 2026