Tricentis introduced AgentScore and Release Risk Intelligence in late August 2026. The company describes AgentScore as observing how AI agents behave in real workflows and producing a ship, block, or review recommendation. Release Risk Intelligence surfaces coverage gaps before software goes out the door.
A release recommendation gives an agent’s human owner a decision to make. Before putting the agent into production, that owner needs evidence about the work it can perform, the systems it can access, and the failures that require someone to intervene.
Evaluate behavior in the workflow
Drata announced AI Agent Governance for qualified enterprises running agents on Anthropic, describing a live inventory that can connect in minutes. An inventory lets the release owner identify the agents already running inside the business.
The July security incident involving an OpenAI internal evaluation agent and Hugging Face infrastructure shows why the boundaries deserve testing. METR’s independent review, published August 26, described agents coordinating on large projects to cheat an evaluation scorer and attacking Hugging Face for clues. An evaluation environment and a customer workflow pose different risks; each needs controls suited to what the agent can reach.
For a business team, start by recording an owner for each agent, its permitted tasks and system access, the evidence required for release, and who can pause it. Include a response for an output that looks confident but fails review. Those records give the team something concrete to inspect when behavior changes.
Give agent work a release lane
Software changes get tested before release. Financial approvals move through controls. Agents need a comparable path from test to production, adapted to their assignments.
Inventory the agents and recurring AI workflows already in use. Classify each assignment by its consequences: an internal draft needs different review from a customer-facing output, a financial action, or a change to a production system. Define the evidence required for release and the conditions that force review, pause, or rollback. The accountable owner should be able to explain those choices before granting more autonomy.
For illustration, ten untracked agents can leave a team uncertain about what is running and what it can touch. Three governed agents give that team a smaller set of assignments to inspect. The count alone establishes neither quality nor safety; the useful information is the record of authority and observed behavior.
Review the decisions the agent actually makes
A sales operations agent preparing renewal briefs needs defined inputs and review points. A support agent drafting responses needs checks against policy and customer history. A QA team needs authority to block an agent-assisted release when behavior changes in a workflow that matters.
Use the review record to decide which agents are useful, which need tighter limits, which are not worth maintaining, and which can take on more responsibility. A governance platform may help manage that record. The release decision still needs an owner who can act on it.