Fetch.ai launched AEVS, short for Agent Execution Verification System, with a public release dated May 12, 2026, according to the visible PyPI listing for aevs 0.2.0 and the GitHub release record for fetchai/AEVS-sdk. The project is presented as an open-source Python SDK for creating verifiable receipts for AI agent tool calls, rather than as a new agent framework. The visible references are GitHub fetchai/AEVS-sdk, PyPI package aevs, and the project site aevs.fetch.ai.
AEVS as a Verification Layer, not an Agent Framework
Fetch.ai claims AEVS provides a receipt for every agent action, including examples such as sending an email, charging a card, updating a record, approving a payment or processing a refund. The company also says the integration can be added with minimal code and without changing existing tools. What is independently visible in the public materials is that the SDK supports Python 3.10+, provides adapters for LangChain, LangGraph and MCP, and can be installed through PyPI using pip install aevs.
AEVS works by sitting between an agent and its tools, capturing the tool input, output, runtime and errors when a tool call occurs. The SDK then builds a receipt, signs it with HMAC, links it to the previous receipt through a hash chain and sends it to the AEVS backend, while buffering receipts locally in encrypted SQLite if needed. This means later verification can check whether the receipt signature is valid and whether the receipt chain remains intact.
Tamper-evident Receipts, but Not an Independent Audit
Fetch.ai describes the system as a verification layer for autonomous agent workflows, useful when agent actions need to be auditable after execution. The project site says a receipt ID can be checked through a public explorer or verification endpoint, while the GitHub documentation shows a public GET /v1/receipts/verify/{reference_id} endpoint that returns fields such as verified, reference_id, receipt_id, tool_name, agent_id, status and input/output hashes.
However, the public material should be framed carefully. Fetch.ai’s marketing language describes AEVS as creating receipts that “can’t be faked” and as “tamper-proof,” but the GitHub threat model is narrower: it says AEVS is tamper-evident, not tamper-proof, and helps detect modification or reordering of receipts after the fact rather than securing a fully compromised host process. No independent technical audit, third-party benchmark or broader rollout timeline is visible in the reviewed material.