At 9:17 a.m. on Oct. 29 (Coordinated Universal Time), version 2.17.0 of the open-source package vaara was released. The tool aims to leave a verifiable execution receipt each time an AI agent calls an external tool. On the same day, AWS introduced a method in which agents approve their own tasks and seek human confirmation only for risky actions.
An agent is an AI program that takes a single goal and handles multiple steps of work on its own. Until now, a person had to click to confirm at every step, so work often stalled, but under the AWS method that confirmation remains only for risky actions. Work gets faster. Which actions are classified as risky is something AWS still needs to explain further.
In the same week, a case emerged in which an AI agent was reportedly used in a wide-ranging hacking campaign targeting financial companies, and an attempt to run the agent in question directly was also made public.
The ability to finish work on its own can be used by attackers too.
vaara has positioned itself as a response to the EU AI Act (the European Union's artificial intelligence law). An agent's tool calls are executed only after passing a predefined policy. Audit records use a hash chain.
The method links each line of the record to the digital fingerprint of the line before it, so that changing even one line in the middle breaks the chain. The developer said recorded times are anchored externally so dates cannot be rewritten later, and that each MCP (a common protocol that lets agents call outside tools) call comes with a proof that a third party can verify independently.
All of these are features described by the developer itself. Whether they hold up in real attack situations and whether they meet the level required by European regulations can only be determined through external verification.
On Oct. 31, a column on trust in medical AI examined hospital environments where multiple agents and tools run together. It argued for setting out which tools may be used and which matters require human approval, and for building a foundation to secure accountability and records.
Agent-related news followed one after another in cloud, finance and healthcare within a single week. As much as how smart an agent is, when it stops to call a person and whether that moment is recorded so it cannot be erased are likely to decide whether it is adopted. These are items that rarely show up on performance charts.
Before long, the inboxes of finance staff will fill up with more receipts for work agents have finished than approval requests. Only a few of them will arrive flagged for confirmation. Staff who spent their days pressing approval buttons will spend them picking out and reading receipts.
