Astra's classifiers halt unauthorized agent tasks outright, Teneo roundup says
Teneo compares Astra's classifier checks on reasoning and actions, which end a task when it oversteps, with Binance's revocable agent permissions and CrowdStrike's scoped approvals.
Agents can transact. Now product teams are defining where they must stop. - OpenAI released GPT-6 Astra with classifiers that inspect reasoning and actions for unauthorised behaviour. Those checks can stop a task. In the API, a stopped run ends instead of continuing through a confirmation flow. - Binance put revocable permissions beside Agent OS capabilities spanning trading, market data, wallets, and payments. Its own Codex example still surfaced each spot order for approval. - CrowdStrike announced Falcon for the Anthropic Claude Marketplace with scoped permissions, human approvals, and auditability built into the security workflow. The pattern is practical: capability is no longer enough to describe an agent product. Buyers and builders also need to know which action is allowed, what requires approval, and what happens when a limit is reached. A receipt can record an unwanted action perfectly. It cannot decide whether that action should run. Teneo gives builders a concrete unit for setting that boundary. ProcessTask defines command behaviour, builders publish command pricing, users see the price before execution, and payment runs through x402. Changing an agent's commands or capabilities also resets its public review status. The next useful test is a stopped task: set the permitted command and task budget, count retries against it, then verify the workflow stops before another paid call. Read the full round-up 👉: https://teneo-protocol.ai/blog/agents-got-receipts-now-they-need-limits