Agent mesh RFP checklist: 40 questions for vendors

0 of 40 checked

A. Interoperability and protocols (1-7)

1. Which A2A versions do you support (v0.3, v1.0, v1.0.1), and in which direction (client, server, both)?Why it matters: v1.0.0 introduced breaking changes in March 2026.Vendor answer: ____________________________________________________________
2. How do you support MCP: as a client, a server, or both, and which spec revision?Why it matters: MCP revisions add features such as Tasks for long-running work.Vendor answer: ____________________________________________________________
3. Can agents built on other frameworks (for example LangGraph, CrewAI, Google ADK) participate without code changes?Why it matters: Rewriting agents to join a platform defeats the purpose.Vendor answer: ____________________________________________________________
4. Which cloud agent runtimes can you connect (Microsoft Foundry, Amazon Bedrock AgentCore, Gemini Enterprise)?Why it matters: Most estates span more than one cloud.Vendor answer: ____________________________________________________________
5. Which SaaS agent platforms can you connect (Salesforce Agentforce, SAP Joule, Workday Sana, Copilot Studio, ServiceNow)?Why it matters: Business agents live close to business data.Vendor answer: ____________________________________________________________
6. Which SDKs and languages do you provide?Why it matters: Internal teams need to connect custom agents.Vendor answer: ____________________________________________________________
7. How are agents discovered and registered, and can that registry sync with others we run?Why it matters: Duplicate inventories drift.Vendor answer: ____________________________________________________________

B. Coordination and reliability (8-14)

8. How is work routed to a specific agent?Why it matters: Ambiguous routing causes duplicated or dropped work.Vendor answer: ____________________________________________________________
9. Do you track delivery and processing status per agent, per message?Why it matters: You need to know which agent failed, not just that something failed.Vendor answer: ____________________________________________________________
10. What prevents two agents from replying to each other indefinitely?Why it matters: Agent loops burn tokens and money.Vendor answer: ____________________________________________________________
11. Do you guarantee exactly-once processing across distributed agents?Why it matters: Duplicate processing creates duplicate actions.Vendor answer: ____________________________________________________________
12. What happens when an agent crashes mid-task, and how does it catch up?Why it matters: Long-running work will see failures.Vendor answer: ____________________________________________________________
13. Is there a central orchestrator, and what happens if it is unavailable?Why it matters: A single controller is a single point of failure.Vendor answer: ____________________________________________________________
14. How many agents and concurrent tasks have you documented limits for, and what are they?Why it matters: Published limits make capacity planning possible.Vendor answer: ____________________________________________________________

C. Shared context and memory (15-20)

15. How do agents from different vendors share conversation history and task state?Why it matters: Lost context is the most common handoff failure.Vendor answer: ____________________________________________________________
16. Is there shared memory across agents, and who can read or write it?Why it matters: Memory without access control leaks data between tasks.Vendor answer: ____________________________________________________________
17. Can work be scoped to a task, with only relevant agents and people included?Why it matters: Scoping limits noise and exposure.Vendor answer: ____________________________________________________________
18. How long is context retained, and can retention be configured?Why it matters: Retention affects both cost and compliance.Vendor answer: ____________________________________________________________
19. Can we export conversation history, memory and audit data?Why it matters: Exit and e-discovery require exports.Vendor answer: ____________________________________________________________
20. How is context passed when a task moves from one platform's agent to another's?Why it matters: This is where point-to-point integrations usually break.Vendor answer: ____________________________________________________________

D. Human participation (21-26)

21. Where do people see agent work: in the same place agents coordinate, or in a separate console?Why it matters: Separate tools create blind spots.Vendor answer: ____________________________________________________________
22. Can a person address an agent directly and redirect its work?Why it matters: Supervision needs a way to intervene.Vendor answer: ____________________________________________________________
23. Which actions can require human approval, and how is approval recorded?Why it matters: Approvals must be auditable.Vendor answer: ____________________________________________________________
24. Can a person override an agent's decision, and is the override visible to other agents?Why it matters: Agents should not keep working from a decision a person reversed.Vendor answer: ____________________________________________________________
25. Do you integrate with Slack or Microsoft Teams?Why it matters: People work where they already communicate.Vendor answer: ____________________________________________________________
26. How many people and agents can take part in one task or room?Why it matters: Cross-functional work involves many participants.Vendor answer: ____________________________________________________________

E. Governance, identity and audit (27-33)

27. How does each agent get an identity, and how does it map to our identity provider?Why it matters: Agents need identities as distinct as employees'.Vendor answer: ____________________________________________________________
28. What RBAC model applies to agents and people?Why it matters: Least privilege applies to agents too.Vendor answer: ____________________________________________________________
29. Can you show which agent did what, when, and under whose authority, including the full delegation chain?Why it matters: Delegated actions must be traceable to a person.Vendor answer: ____________________________________________________________
30. What isolation tiers exist between personal, team and organization-wide resources?Why it matters: Isolation limits the impact of misconfiguration.Vendor answer: ____________________________________________________________
31. How are credentials and API keys stored and rotated?Why it matters: Agent credentials are high-value secrets.Vendor answer: ____________________________________________________________
32. Can policies be enforced at runtime, not only reviewed after the fact?Why it matters: After-the-fact audit does not stop a bad action.Vendor answer: ____________________________________________________________
33. Which compliance attestations and reports can you provide under NDA?Why it matters: Procurement will ask; get documents, not claims.Vendor answer: ____________________________________________________________

F. Deployment, operations and commercial (34-40)

34. Which deployment models do you offer: SaaS, private cloud, on-premises, hybrid?Why it matters: Data residency and network rules vary by workload.Vendor answer: ____________________________________________________________
35. Which clouds and regions do you run in?Why it matters: Latency and residency depend on region.Vendor answer: ____________________________________________________________
36. What observability do you expose (traces, metrics, logs), and in what formats?Why it matters: Operations teams need to see agent behaviour in existing tools.Vendor answer: ____________________________________________________________
37. How does pricing scale: per seat, per agent, per message, per execution, or per compute hour?Why it matters: Agent traffic grows faster than headcount.Vendor answer: ____________________________________________________________
38. Is there a free or low-cost tier for a pilot, and what are its limits?Why it matters: A pilot should run on the real product.Vendor answer: ____________________________________________________________
39. What support and SLA terms apply at each tier?Why it matters: Production agents need response commitments.Vendor answer: ____________________________________________________________
40. What does it take to leave: data export formats, notice periods, and agent portability?Why it matters: Agents connected via open protocols should stay portable.Vendor answer: ____________________________________________________________

How to use the checklist

Send groups A to D to collaboration-layer candidates, group E to all vendors, and group F to all vendors. Score each answer 0 (no), 1 (partial or roadmap) or 2 (documented and demonstrated), and weight groups to match your priorities. The rubric we use for the mesh fit score is on the About page.

Frequently asked questions

Can I download this checklist?

Use the Print / Save as PDF button. The print layout includes answer lines for each question.

Which questions matter most?

For cross-vendor estates, questions 3, 9, 10, 15, 21 and 29 separate collaboration layers from gateways and runtimes.