Dual-Surface Architecture: Serving Humans and Agents from the Same Tool Layer

Official Schedule Context

Every enterprise AI talk right now is about capability. Almost none are about containment. That's

the gap this talk fills, because it's where regulated deployments actually die. The Deterministic

Harness is the set of rigid rails around a model: schemas, data contracts, tool boundaries, and

audit paths. These rails are what turn a probabilistic model into a deployable enterprise asset. The

idea isn't new. Aviation wraps pilots in envelope protection. Nuclear wraps reactors in passive

safety. Banking wraps algorithmic trading in transaction limits. Every regulated industry figured

out the same thing eventually: high-variance systems only become deployable when wrapped in low-

variance containment. Enterprise AI is catching up, not inventing. I'll walk through the single

governed MCP and API server we built at Carlyle, and the architectural decisions behind it. You'll

leave with four things: 1. A phased rollout model where each phase earns the next. Moving from

locked-down reads to trusted writes isn't risk mitigation. It's trust compounding. Each phase

generates the observability that underwrites the autonomy granted in the next one. Skip a phase and

you don't save time. You destroy the evidence base that would have justified the next step. 2. One

contract, two surfaces. A single data layer that serves both the human UI and the agent. The

institution then has exactly one answer to any question either might ask. When the agent and the UI

disagree, users lose trust in both. 3. An intent based feedback loop that captures what LLM

providers structurally cannot. The gap between what users tried to accomplish and what the system

actually delivered is invisible to Anthropic, OpenAI, and Google. Only the harness owner sees it. We

close that loop back into the governed server, and it compounds into differentiation that model

providers cannot replicate from where they sit. 4. The failure modes we hit and what we'd redesign.

A pre mortem folks will inherit for free, from two regulated industries where a wrong answer has a

named owner.

No official session recording transcript was found by exact title match on the AI Engineer YouTube channel during this run.

Pending transcript synthesis when an official recording or confirmed matching video is available.