Amazon hosting OpenAI's flagship models on its own infrastructure is Amazon admitting that owning the best model was never really the plan, owning the harness customers run any model through was. Every enterprise that picks Bedrock now gets a choice between OpenAI, Anthropic, and Amazon's own models inside one billing relationship, one security posture, and one API surface, which makes the model itself a swappable component rather than a platform lock-in decision. That same roundup listing five different open agent protocols side by side in one guide is the tell for where the real complexity has moved: nobody is trying to pick the winning protocol anymore, they are trying to help developers use several at once without caring which one wins.

AWS just demonstrated the component position from the inside: by hosting a rival's model as one interchangeable option among several, Bedrock keeps the accountability boundary, security, billing, compliance, observability, at the harness layer regardless of which lab's model handles a given request. That is the deliberate answer to single-orchestrator dependence, and it is exactly why the harness, not the model underneath it, is where FuzeBox AEOS and platforms like it are built to sit.

Salesforce just removed the price tag from the exact capability that lets a customer catch an agent behaving badly before it becomes an incident, which only makes sense if the company has concluded that unmonitored agents are a bigger threat to Agentforce's adoption curve than the lost metering revenue. Consolidating trust, health, and quality metrics that used to sit in separate dashboards is the more telling architectural move: it means Salesforce is treating "is this agent trustworthy" as one continuous signal instead of three disconnected reports a team has to reconcile manually. Every vendor selling multi-agent systems is watching the same failure pattern this brief has covered repeatedly, an agent doing something individually defensible that becomes a problem only once you see the full chain, and a waterfall trace that follows execution across subagents is the direct structural answer to exactly that blind spot.

Agent Script is Salesforce's answer to a problem every enterprise running agents has hit by now, that letting a language model reason freely through every step of a regulated process is a compliance nightmare even when the model gets the answer right, because "usually correct" is not the same as "provably follows the rule every time." By making some steps deterministic and only letting the model reason freely where that flexibility is actually valuable, Salesforce is drawing the exact line this brief keeps returning to: which decisions get automated with judgment, and which get locked down as a rule the system must never deviate from. Retiring new-agent creation on the old builder without killing existing agents is a sensible migration pattern, but it also means every Salesforce customer building agents from this week forward inherits Agent Script's rules whether or not they asked for that architectural opinion, which is worth knowing before you start your next build.

HR and finance are two of the functions where a wrong or biased automated decision creates the fastest path to a lawsuit or a regulatory inquiry, which makes them a deliberately hard place for Workday to stake its credibility, not an easy one. Betting an entire flagship event on those two functions being ready for agentic deployment is Workday signaling that it believes the governance and audit-trail problem, not the underlying model capability, is now solved enough to put in front of a room of practitioners who will ask pointed questions about exactly that. The market lesson generalizes past Workday: the vertical-function ERP vendors are all racing to build their own agent layer inside HR, finance, supply chain, and CRM rather than ceding that layer to a horizontal orchestrator, which is going to leave enterprises juggling several vendor-specific agent architectures instead of one unified one for the foreseeable future.

A security agent that acts autonomously on bad data is not a marginally worse security agent, it is a liability that moves at machine speed, which is exactly why "clean up your data first" has quietly become one of the most fundable pitches in enterprise AI right now. Beacon's bet is that most enterprises still cannot answer a basic question, whether the telemetry feeding a security agent is complete, current, and consistent across every vendor product it touches, and that gap is what keeps security teams manually reviewing agent output instead of trusting it to act. The 300% revenue growth figure, if it holds, says enterprises are already paying to solve this before regulators or auditors force the question, which puts data trust ahead of model capability as this quarter's actual bottleneck in security automation.

The learning-authority dilemma runs through this story from the data layer up: a security agent's decision ability to act, isolate a host, revoke a credential, block traffic, should never exceed the formal authority a human is willing to grant it, and that authority should scale with how trustworthy the underlying data actually is, not with how confident the agent sounds. Beacon is selling exactly the missing piece that lets an organization widen an agent's authority safely, a verified, consistent data foundation, rather than widening it on faith and finding out where the rule debt was hiding after an incident.