Agent-Gantry & ADAPT-Agent
On this page
The problem
Giving an agent more tools makes discovery and execution harder to reason about. A useful tool description needs to become a provider-compatible schema, a selected action and a controlled execution with observable results.
What I’m building
Agent-Gantry is a Python tool orchestration library with semantic selection, provider schema conversion, controlled execution and MCP / A2A integration. ADAPT-Agent explores policy controls, adversarial detection, taint tracking, evaluation and framework adapters.
Design decisions
Selection and execution are distinct steps. A tool being relevant does not establish that an action is permitted. Policy and data provenance need to survive across calls and framework boundaries.
MCP and A2A support make interoperability practical, while telemetry gives a way to investigate what actually ran.
Evidence and status
Both public repositories include implementation and usage material. The projects explore different parts of the agent infrastructure stack. Their usage material covers the supported interfaces and policy controls.
Explore ADAPT-Agent for its policy and evaluation work.