A running log of agent frameworks put under real bounds: a spend cap, a capability ceiling, and a signed receipt for every decision, added through @agentguard-run/mcp with no changes to the framework. Each post is a setup you can reproduce in about ten minutes, and every verdict in it comes from the published SDK, not a mockup.
Add AgentGuard as a Goose MCP extension, give the agent a five dollar cap and a read-only ceiling, and watch it get refused when it reaches past either one. Receipts verify offline.
buzz-acp harnessBlock's Buzz built authorization and authorship: scoped grants, delegation, revocation, accounting. What it does not carry is enforcement on the economics of an action. Wire AgentGuard into the buzz-acp harness through one env var and that seam closes, with zero Buzz-side code.
Every post answers the same question the same way, so a new one is a short job, not a project:
1. What the framework does, and the exact bound it is missing. 2. The real setup: install @agentguard-run/mcp, point the framework at it, set a policy. 3. Three runs: one allowed with a receipt, one refused by the spend cap, one refused by the capability ceiling. 4. Verify a receipt offline, then tamper with one and watch verification fail. No invented numbers, no logos we do not have. If a framework cannot be bound cleanly, the post says so.