Agents are leaving the chat window. Three layers your platform now needs
Last week, a run of announcements showed AI agents moving out of chat and into the plumbing of enterprise platforms. Look past the product names and three architectural layers keep appearing: a way to connect, a safe place to act, and a data layer that can keep up.
1. A connection layer built on open protocols
- At SAP TechEd (30 October), SAP added an agent gateway for agent-to-agent interoperability and an AI agent hub to discover, document and check SAP and non-SAP agents for compliance.
- GitHub’s Copilot now carries cloud agent sessions across web, mobile, CLI, VS Code and Slack/Teams, and previewed “connect-once” connectors for around 26 services.
- Android Studio’s “Bring Your Own Agent” lets developers plug in Claude, Codex, Antigravity or any agent that speaks the Agent Client Protocol.
Architect’s note: Agents are becoming interchangeable, so lock-in moves from the model to the integration. Standardise on open protocols (MCP, ACP, agent-to-agent) and keep one registry of every agent, what it connects to and who owns it.
2. A sandbox for what agents do
On 27 October, Akamai launched Agent Sandbox in beta. It runs code written by agents in WebAssembly isolation, separate from the agent’s own runtime, with limits on compute, memory, outbound connections and system capabilities.
Architect’s note: Treat agent-generated code like untrusted input, because it is. Run it in an isolated environment with no standing credentials, allow outbound calls by explicit list only, and log every execution.
3. A data layer built for bursty, unplanned queries
At Cloudera EVOLVE26, PhoenixAI (CelerData) argued that agents generate large bursts of unplanned queries, often across several databases at once, and still expect real-time answers. Its answer was a context environment that limits agents to the data they actually need.
Architect’s note: Capacity plans built for human dashboards will not hold. Put agents behind a governed data access layer with per-agent quotas, caching and cost attribution, and load-test it with agent-style traffic before go-live.
The takeaway
The model is the easy part to swap. The lasting value, and the lasting risk, sits in how agents connect, where they execute and what data they can reach. Design those three layers deliberately, and the agents you choose next year will slot straight in.
At KSL, we help organisations design agent platforms with clear protocols, safe execution and governed data access. If your agent pilots are about to meet production traffic, let’s talk.
Sources
- SAP beefs up Business AI Platform and Joule to drive agentic AI adoption (ITWeb)
- AI Agents News, week of 29 October 2026 (AI Agent Store)
- Build your way: use any AI agent in Android Studio (Android Developers)
- Akamai launches AI inference service and agent execution environment (Akamai)
- Agentic AI architecture demands real-time databases (SiliconANGLE)