Every benchmark now admits it: the harness changes the score. Yet most teams are quietly locking themselves into a single agent runtime — its config format, its skill system, its context conventions — the same way we once locked into a CI vendor.

Haggai works as a consultant across multiple client environments, each mandating a different runtime: Cursor here, Claude Code there, open-source harnesses elsewhere. That constraint forced a discipline: treat the harness as a replaceable execution engine, and invest in what's portable — versioned skills, layered context schemas, and the feedback loop the agent runs inside (what he calls loop engineering, the successor to prompt and context engineering).

This session walks through a working setup: a monorepo of versioned skills and project definitions, an MCP hub running on Kubernetes serving tools to any harness, and the token economics that decide whether an agentic workflow is viable at all. Attendees leave with a concrete portability checklist — what to standardize, what to abstract, and how to evaluate the next harness without rebuilding everything — plus lessons from real client migrations.

Target audience: Developers, DevOps/Platform Engineers, tech leads, and engineering managers adopting AI coding agents — anyone evaluating (or already regretting) an agent runtime choice. Mid-level and up; assumes basic familiarity with AI coding assistants, no specific runtime experience required.
Track fit: AI & ML Integration; also relevant to Platform Engineering and DevOps and Continuous Integration tracks.
Preferred duration: 40–45 min talk (adaptable to 30 min)
Format: Slides + live demo — the same versioned skill executed on two runtimes (e.g., Claude Code and an open-source harness), backed by an MCP hub on Kubernetes. Recorded fallback available.
Technical requirements: Standard A/V (HDMI, mic); internet connection preferred for the live demo, not mandatory.
First public delivery: Yes — this session has not been delivered publicly. It builds on the speaker's blog series on loop engineering and harness portability (portfolio.hagzag.com) and prior Agentic AI in DevOps talks (Reversim Summit, DevOps Days).
Language: English.
Content notes: Vendor-neutral by design — the whole point is not marrying one vendor. All tooling shown is open source or commercially available (Claude Code, OpenCode/Cursor, MCP, Toolhive, Kubernetes); client examples anonymized.