Dynatrace paid $915 million for Arize and classified agents as applications to be traced and scored. Fortinet's purchase of Virtue AI classified them as attack surface. Okta buying Permiso made them non-human identities that need governing after they authenticate. Each deal makes sense on its own. Read together, they're doing something less obvious: every acquisition bakes in a theory about what an agent is, and that theory decides which problems the product will catch for you and which ones it won't. Choosing agent infrastructure right now means choosing a frame, whether or not you meant to.

Dynatrace paid $915 million for Arize and classified agents as applications to be traced and scored. Fortinet's purchase of Virtue AI classified them as attack surface. Okta buying Permiso made them non-human identities that need governing after they authenticate. Each deal makes sense on its own. Read together, they're doing something less obvious: every acquisition bakes in a theory about what an agent is, and that theory decides which problems the product will catch for you and which ones it won't. Choosing agent infrastructure right now means choosing a frame, whether or not you meant to.

A busy week across the agent ecosystem, with movement on several fronts at once. Google's A2A protocol shifted to neutral foundation governance. Two acquisitions pulled observability and security startups into larger enterprise platforms. A new model shipped with introductory pricing that expires in four months.

Meanwhile, questions about oversight arrived from different directions: Anthropic published research on what happens when agents coordinate without supervision, and the UK's legal regulator responded to what happens when professionals skip reviewing AI output before it reaches a courtroom. Here's the full rundown.

A busy week across the agent ecosystem, with movement on several fronts at once. Google's A2A protocol shifted to neutral foundation governance. Two acquisitions pulled observability and security startups into larger enterprise platforms. A new model shipped with introductory pricing that expires in four months.

Meanwhile, questions about oversight arrived from different directions: Anthropic published research on what happens when agents coordinate without supervision, and the UK's legal regulator responded to what happens when professionals skip reviewing AI output before it reaches a courtroom. Here's the full rundown.

Google launched Gemini 3.7 Flash on August 13 at $0.75 per million input tokens and $3.75 per million output tokens, with both rates doubling on January 1, 2027. The model is positioned as Google's "workhorse for coding and agents" and already powers its own agent infrastructure by default.

The promotional window looks like a discount. It functions as something closer to a subsidy for adoption during the period when architectural choices get made. Agent workflows involve calibrating loop lengths, retry budgets, evaluation frequency, and the scope of what gets automated against a cost assumption. Once those calibrations are set and running in production, they tend to stay. Developers don't casually redesign a working pipeline because the price of inference changed.

Google reinforced this by applying the same introductory rate to 3.6 Flash, eliminating any price difference between model generations during the window. You migrate to 3.7 because it's more capable, not because it's cheaper. The capability preference becomes the production default. The production default carries the doubled rate into 2027.

The question practitioners should be asking is whether cheaper inference during this window actually reduces their total cost, or whether it funds more ambitious automation that meets the permanent rate with greater exposure. Promotional pricing shapes what people build. What people build determines what they're committed to when the promotion ends.

Introductory rates: $0.75/M input, $3.75/M output through Dec 31, 2026. Both double to $1.50/$7.50 on Jan 1, 2027.

Predecessor pricing: Google applied the same promo rate to 3.6 Flash, so both models cost the same during the window. Migration becomes a capability decision, not a cost one.

Thinking tokens count as output. 3.7 Flash offers three reasoning levels (low, medium, high); higher settings increase output volume substantially, pushing agent-loop costs well beyond headline input pricing.

Promotional windows aren't synchronized. Claude Sonnet 5's introductory pricing expired Aug 31. Gemini's runs through Dec 31. GPT-5.6 Luna lists lower base rates without a published expiration. Teams choosing providers during one window may face different economics when another closes.

Cost per task vs. cost per token. For agent workflows, the meaningful unit is cost per completed task, including retries, tool calls, evaluation loops, and recovery. Token price is one input to that calculation.