Prasad Kavuri

Agent Platform Layer

Enterprise Agent Runtime

A good prompt gets an agent working once. A runtime is what keeps it working — safely, reversibly, and observably — as it changes, scales, and touches production data. This is why agent runtime matters more than the prompt itself.

Part of the Enterprise Capability Lifecycle— this page covers Versioning & Runtime.

Why this, not just prompts

Prompt engineering answers "what should the agent say." Runtime engineering answers "what happens when it's wrong, when it needs to change, when it acts on real data, and when ten teams need different behavior from the same agent." Production-managed-agent patterns from industry (prompt registries, canary rollout, session-scoped overrides) point at the same conclusion: the runtime around the model is where enterprise reliability is actually won.

Agent Identity

Agents need their own identity lifecycle — anonymous by default, upgradeable to a verified, Bearer-authenticated identity once trust is established.

Why it matters

Without agent identity, every action an agent takes is unattributable — no audit trail, no permission scoping, no accountability.

Prompt Versioning & Registry

Prompts are versioned artifacts with hashes, not inline strings — each version tracked from draft through canary to stable, with an eval score at every promotion gate.

Why it matters

Prompts are the part of an agent most likely to change in production; versioning them like code is what makes changes reviewable and reversible.

Canary Deployment & Rollback

New prompt or policy versions roll out to a small traffic percentage first. If the eval score drops below a gate during the canary window, rollback is automatic.

Why it matters

Removes the choice between "ship fast" and "ship safe" — canary plus an automatic eval gate gives you both.

Session Overrides & Policy Enforcement

Per-team, per-session overrides (model tier, temperature, tool access, max tokens) let policy adjust agent behavior without touching the underlying prompt or code.

Why it matters

Compliance and cost constraints differ by team; overrides let one agent runtime serve all of them without forking the codebase.

Tool Registry

Tools are registered with declared permissions, availability, latency, and cost-efficiency metrics — not wired ad hoc into agent code.

Why it matters

A registry is what turns "the agent can call some functions" into a governed, auditable capability surface.

Capability Registry

The Tool Registry, Prompt Registry (Agent Lifecycle tab), and the application-level Skills Catalog are three facets of one logical layer: everything an agent runtime can call on, versioned and governed rather than wired in ad hoc. A Workflow Library / Enterprise Playbooks tier is a natural extension of this layer that this portfolio does not yet implement.

Why it matters

Enterprises rarely lack capabilities — they lack a single place that knows which ones exist, who owns them, and what depends on what. Naming the registry as one layer, even when its parts live on different pages, is what makes that discoverable.

Human Approval (HITL)

Higher-risk agent actions pause at an explicit checkpoint for human approval before executing, with the decision logged to an audit trail.

Why it matters

The gate between an agent that reasons and an agent that acts unsupervised in production — required for anything touching money, customer data, or irreversible actions.

Memory

Short-term session memory, long-term retrieval memory, and episodic session replay each serve a different purpose and need different retention and access rules.

Why it matters

Undifferentiated memory is a data-governance liability; scoped memory tiers are what make retention policy and PII handling tractable.

Observability

Every tool call, connector call, and approval decision emits a structured event with trace ID, duration, and token cost — streamed to a live feed.

Why it matters

When an agent misbehaves in production, observability is the difference between a five-minute root cause and a multi-day investigation.