Agent Only MCP: shared verified troubleshooting memory for AutoGen teams

[vegelate (vegelate)]

Maintainer disclosure: I maintain Agent Only, so this is a self-authored project introduction based on my own multi-agent development workflow.

The problem I kept hitting was repeated diagnosis. One coding agent would discover why a build, deployment, or tool call failed, but the next session had no durable record and repeated the same dead ends. Agent Only provides shared troubleshooting memory over MCP: search earlier diagnoses, verify them against the current repository, and save only a confirmed resolution.

For AutoGen, use McpWorkbench with StdioServerParams and pass the command plus argument list. The local server values are:

command: npx
args: -y, agent-only-mcp

The shortest standalone launch command is:

npx -y agent-only-mcp

Node.js 20 or newer is required. The stdio bridge runs locally but connects to the hosted Agent Only service, so it is not an offline-only integration. A safe workflow is read/search first, independently verify the advice, and request explicit approval before writing. Do not store credentials, API keys, proprietary source code, customer data, or personal information; keep tool permissions narrowly scoped.

Source and setup notes: https://github.com/vegelate/agent-only-mcp

[ANDERSON AYORA (Ayorinha)]

I think verified troubleshooting memory should be treated as a controlled knowledge store rather than simply shared agent context.

A useful flow would be:

"Agent observation → Evidence validation → Verification → Memory write → Future retrieval"

Before a troubleshooting result becomes persistent memory, I would record its provenance: which agent produced it, which tool or MCP server was involved, the relevant evidence, timestamp, and the state/version of the system.

I would also separate the confidence of the information from its authorization to become memory. An agent may propose a finding, but persistent memory should require explicit validation rules.

For MCP-based systems, I would additionally verify that the tool result actually supports the stored conclusion. This helps prevent a compromised or misleading tool response from becoming persistent knowledge and influencing future agents.

For high-risk memory writes, a deterministic policy could return:

"VERIFIED / SUPPORTED / PARTIALLY_SUPPORTED / CONFLICTING / UNVERIFIED"

rather than forcing every observation into a binary trusted/untrusted state.

Finally, I would keep an append-only audit trail and make memory updates replayable. If a stored troubleshooting fact is later found to be incorrect, the system should be able to identify exactly which evidence caused it to enter memory and which downstream decisions relied on it.

That gives shared agent memory a much stronger integrity model while still allowing agents to collaborate through MCP.