We connected A2A delegation to MCP tool execution—and tested where stale evidence gets stopped
FreshCtx 0.14.0 launch package
Indie Hackers
New-post title
We connected A2A delegation to MCP tool execution—and tested where stale evidence gets stopped
Post
An AI agent reads a customer record, policy, approval, balance, or deployment state and makes a reasonable decision.
It then delegates the task to another agent.
That agent delegates the final operation to an MCP tool.
What happens if the original evidence changes somewhere between those steps?
Most systems can tell you which agents exchanged messages. That does not necessarily establish whether the evidence behind the final action was still current.
That is the boundary we added to FreshCtx 0.14.0.
FreshCtx now supports signed, time-bounded evidence references across Agent2Agent delegation. The receiving agent verifies the delegation record and revalidates the declared evidence before its executor starts.
We built a three-agent demonstration that ends at a real MCP tools/call boundary:
Current evidence: the final MCP tool executes exactly once.
Changed evidence: the delegated action is blocked.
Evidence that cannot be checked: the action is blocked.
Expired, tampered, missing, or wrong-recipient delegation: the receiving executor never starts.
Unrelated evidence changes: the declared action can continue.
The delegation record carries identifiers, expiry, chain links, and evidence references. It does not carry prompts, credentials, source contents, tool arguments, or business payloads.
The scope is deliberately precise. FreshCtx does not decide whether the original evidence was true, whether the agent selected the right source, or whether the user was authorized. It checks whether the declared evidence behind this action is still current and whether the delegation record arrived intact.
A2A gives agents a standard way to discover and delegate to each other. MCP gives agents a standard way to reach tools. FreshCtx now places one consistent evidence boundary across that path.
The release is public:
GitHub: https://github.com/Hyperwise-LLC/freshctx
PyPI: https://pypi.org/project/freshctx/0.14.0/
A2A documentation: https://github.com/Hyperwise-LLC/freshctx/blob/main/docs/A2A_GUARD.md
Three-agent example: https://github.com/Hyperwise-LLC/freshctx/blob/main/examples/a2a_to_mcp_delegation.py
Install:
python -m pip install 'freshctx[a2a-mcp]==0.14.0'
I am looking for one specific kind of feedback: if you operate an A2A or MCP workflow that can modify money, access, customer records, infrastructure, bookings, or approvals, run the example and tell me where its evidence chain differs from your real system.
A boring result is useful. A broken assumption is even more useful.
