The agent platform is standardising: protocols, observability, governance
Oracle put MCP and A2A servers inside its database, Dynatrace wired observability into Gemini Enterprise, Stacklok made the case for Kubernetes-native agent platforms, and Workato laid out how to govern the lot.
- Oracle says its Autonomous AI Database now includes managed MCP and A2A servers that are built-in, fully managed and multi-tenant, and the two servers can be enabled independently. Self-reported m:…3379
- Dynatrace says its expanded Google Cloud collaboration brings observability into Gemini Enterprise agent workflows, letting teams trace AI activity and monitor latency, token consumption and costs, with observability data also available in the Google Antigravity CLI. Self-reported m:…4346
- Stacklok says its Mecatl agent harness is Apache-2.0 licensed, enforces deny-dominant permissions and budgets, and writes an audit record for every tool call; Stacklok supports air-gapped installs. Self-reported m:…0221
- Workato advocates a three-part governance pattern: a governed orchestration layer, outcome-based project prioritisation, and platform consolidation; it says Asana reached 86% AI adoption in a first quarter and saw a 10% increase in bookings in that quarter. Self-reported m:…6232
Protocols move into the database
Oracle has put the two agent protocols directly inside its Autonomous AI Database, which now includes managed MCP and A2A servers that put tools and agents defined with the Oracle Select AI Agent Framework within reach of compatible clients. The MCP server lets compatible clients discover and connect to those tools, and the A2A server lets compatible clients discover and invoke the agents. Oracle's shorthand: MCP is the path from a client to a tool, and A2A is the path from one agent or application to another agent. m:…3379
The point of building them in is that there is no separate MCP server infrastructure to deploy. Oracle describes the servers as built-in, fully managed and multi-tenant, and they can be enabled independently: turning one on or off does not change the status of the other. They are switched on through the Tool configuration tab in the OCI Console, and access is managed through identity, authorisation and auditing, together with database access controls. The database is becoming where the agent's tools live, not just where its data sits. m:…3379
Watching agents in production
Dynatrace is extending its Google Cloud collaboration to bring observability, automation and AI intelligence into Gemini Enterprise agent workflows. The company says teams can trace AI activity, monitor latency and token consumption, manage costs and link performance to outcomes. The expansion includes deeper Gemini Enterprise integration, enhanced AI observability for Google Cloud environments and broader availability through Google Cloud Marketplace. m:…4346
Dynatrace for Gemini Enterprise lets teams investigate issues and find root causes using natural language, without switching tools. Arize, a Dynatrace company, lets teams trace AI applications and agents, evaluate outputs and investigate where and why they fail. Dynatrace also provides observability data directly into the Google Antigravity CLI, and is extending it into Antigravity workflows through a plugin in the Antigravity Plugin Marketplace, so developers get insights and root-cause analysis inside the terminal. Dynatrace points to InPost, which it says uses the setup for AI-driven logistics forecasting, workforce planning and shopping agents. m:…4346
A platform, not a framework
Stacklok's argument this week is that authoring agents and running them are two different jobs, and the industry has only solved the first. LangChain and CrewAI are excellent at authoring agents, according to Stacklok; the search for alternatives usually starts at the second job, running, governing and scaling those agents in production. Stacklok describes itself as a Kubernetes-native AI agent platform built around an open agent harness, for enterprises that need to run, govern and scale production agents on their own Kubernetes infrastructure, completely self-hosted. Its Mecatl harness is Apache-2.0 licensed and written in Go; it enforces deny-dominant permissions, approval flows and budgets on turns, tokens and tool calls, and writes an audit record for every tool call. m:…0221
Stacklok surveys the alternatives with a sceptical eye. Kagent is an open-source CNCF Sandbox project that runs agents on Kubernetes in CRD-managed pods, with Postgres-backed sessions, checkpoint and fork, and gVisor or Firecracker sandboxing. TrueFoundry's commercial platform adds LLM, MCP and Agent Gateways, with over 1,000 models across more than 35 providers. Hyperscaler platforms from AWS, Google and Microsoft offer managed runtimes, but the trade-off, Stacklok says, is that the agent stack is tied to the provider's cloud control plane. Stacklok says it is completely self-hosted and documents an air-gap install, so agents, tools and model traffic stay inside the customer's boundary. Its founders, Craig McLuckie and Joe Beda, co-created Kubernetes, the company notes. m:…0221
Governing the builders
Workato's post is the governance manual for the world the other three describe. Its pattern has three parts: put a governed orchestration layer between AI and your data, tie every project to a defined outcome, and consolidate onto fewer platforms. Workato CIO Carter Busse advises CIOs to get in front of MCP sprawl and build the MCP servers themselves, because connecting LLMs to company data changes the way people work. The MCP Registry is a central catalog of an organisation's MCP servers, and the MCP Control Plane governs servers teams compose, prebuilt servers and third-party servers connected by proxy. m:…6232
Each server authenticates users by API token, Workato Identity credentials or single sign-on through an external identity provider like Okta. Tool-level role-based access control lets admins grant a group access to a server and restrict which tools it can call, without standing up duplicate servers. Verified User Access carries the caller's identity through to the underlying app, so backend calls run under that user's own identity and permissions instead of a shared service account. Server-level logs capture the calling user, authentication method, source IP, tool name, inputs and outputs, duration and success or failure, and can stream to a SIEM. m:…6232
Workato backs the pattern with customer stories. SentinelOne prioritises AI projects by specific outcomes, such as renewals automation with a defined outcome and clear return. Asana, the company says, reached 86 percent AI adoption in a quarter and saw a 10 percent increase in bookings in that quarter. Both teams landed on consolidation. Its Model Gateway is a single OpenAI-compatible point of control between applications and LLM providers, with policy enforcement for rate limits, token limits and IP access, and the company says its internal token costs fell by 60 percent after adopting it. Its AIRO tool uses specialised AI agents to build automations and MCP servers from natural language, and is available in Claude Code and Cursor through the AIRO MCP server. Taken together, the week's announcements describe a layer settling into place: standard protocols, observability, platforms, and governance. m:…6232
What is still unresolved
- Whether MCP and A2A become the settled standards or one of them fades as agent frameworks converge.
- How much of the observability and governance tooling teams actually adopt once agents move from pilots to production.
- Whether enterprises accept hyperscaler control planes or insist on self-hosted platforms like Stacklok's.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.