The keynote, "The Hard Part Isn't the Code", frames the developer as an heir: applications, workflows and data shaped by years of change, arriving without the context to say what depends on what. "CRM Case Design" is the same complaint at data-model scale, a Case table bent by every business unit. The enterprise problem being sold to is not generation, it is recovering intent from a running system before changing it.

Our work: Context Surfaces, History as Retrieval Context, Agent Memory as Institutional Knowledge, and the behavioural-rebuild pair Rebuilding Clojure from behavior / Rebuilding from the spec, which ask what a system's observed behaviour lets you reconstruct when the documentation is gone.

"Bring your own harness" names Claude Code, Cursor and Copilot outright, and the SDK ships an AI skill that hands the agent best practices, live instance details and terminal commands for installs and ATF suites. A two-session lab builds an app with Claude Code. The platform has stopped insisting that development happen inside its own IDE; it now competes to be the best-described target for someone else's agent. Build Agent, its own agent, runs alongside, with an "Autonomous Engineer" that turns an ITSM story backlog into a sequenced plan.

Our work: 2026 Q2 Skills (a skill as packaged context, and the sandbox it runs in), CLI Coding Agents: 2026 Q2 Comparison, The Oracle Is the Deliverable. The open question for us is credentials: a vendor skill that reaches a live instance needs a token, and how each harness holds it differs.

The EmployeeWorks lab states it plainly: an assistant is only as trusted as the rules its tools enforce. ToolSmith builds plugins exposed over MCP; an MCP Server Console and "Action Fabric" cover interop "beyond MCP & A2A theory"; a community session argues for specialised assistants sharing tools over one universal assistant. The unit of governance is the tool call, not the model.

Our work: The Agent Integration Layer, September 2026, Agent Permission Guardrails, Agent Identity and Attestation.

Three sessions and an expo game sit on one curve. "AI Agent Sprawl" is the first symptom: duplicated agents, actions and integrations nobody can maintain. "Governing AI Agents at Scale" is the second: hundreds of agents in production, with confidence thresholds, approvals and observability. "Governing Agentic Development End to End" moves governance into App Engine Management Center and AI Control Tower, and the "Super Intelligence Control Tower Challenge" turns an AI centre of excellence into a draft under budget and regulatory pressure. Governance is offered as a product surface, and the pitch is that it speeds teams up.

Our work: Agent Telemetry Systems, Bot and Agent Trust Management Software: The 2026 Landscape, Who Does What, With Agents.

The Flow Designer and Playbooks session is about agentic AI colliding with "the rules your enterprise can't bend": the workflow engine is the fixed frame, and the agent reasons inside it. Two labs make the frame concrete. "Inside a Running Flow" is about data going stale while a flow executes. "Make Your Flows Listen" brings external events in as triggers instead of custom integrations. Enterprise automation is still state machines; the agent is a step, not the controller.

Our work: Every View Is Late (a value read mid-flow is a projection with an age), TLA+ for System Design.

"From Update Sets to Automated Pipelines" puts Git as the source of truth, with automated regression and shorter change review, as the prescribed path for a platform whose native unit of change is still the update set. "AI Agent for ATF Maintenance" names the cost of not getting there: test suites that drift out of coverage after every upgrade, and an agent to find the gaps. Testing is the bottleneck here as it was at the Antithesis evening a week earlier, from the enterprise side.

Our work: Property-testing the crowsnest client (a drift sentinel that caught a real contract change), Property-Based Test Shrink Discipline, Elenctic Vibe Code Review.

Lux, a new UI framework on Lit with an editor-side lab, and the release-specific Studio sessions are platform news rather than a pattern. The expo demos (live IoT plant data, a cocktail fulfilment bot, a React photobooth) show the platform's reach, not a focus area.