The agentic enterprise

Generic agents are the scaffolding. Custom agents are the building.

Why production AI agents need purpose-built design, native tooling and embedded security.

Kevin Keller · kevinkeller.org

Why it matters

Every enterprise is experimenting. Few are in production.

"We're going to see agents in every single part of every single company."

Jensen Huang, GTC 2026

The question is no longer whether, but how. The gap from pilot to production is an architecture problem, not a technology problem.

Three phases

Most pilots stall between demo and production

PhaseWhat happens
Experimentation, 2024 to 2025LLMs call tools, MCP servers appear everywhere
Disillusionment, 2025 to 2026Demos fail in production
Production architecture, 2026 onwardsSpecific agents for specific tasks

Why pilots fail

Four problems appear at production scale

  • Security gaps: credentials per server, not per user
  • Reliability: same prompt, different tool sequences
  • Compliance: audit trails span four systems
  • Cost: round trips, discovery tokens and retries add up

MCP in production

What helps exploration hurts production

  • Dynamic discovery is a liability: capabilities can change without code changes, the "rug pull".
  • Authentication fans out across independent trust boundaries.
  • The audit trail fragments across systems.
  • Protocol overhead grows with thousands of invocations an hour.

The idea

Build specific agents for specific tasks

Not a retreat from the agentic vision. Its maturation.

What works

Five traits of production-grade agents

  1. Purpose-specific design: one job, done well.
  2. Native tool integration: platform APIs, no protocol layer.
  3. Embedded security: own identity, platform RBAC, native audit.
  4. Deterministic tools: a fixed set, defined in code and testable.
  5. Evaluation before deployment: domain test suites gate every release.

Secure by construction

An agent can't run what isn't in its code

  • No Bash, no dynamic discovery: the code is the security boundary.
  • Containers: network and filesystem isolation, resource limits, immutable images.
  • Inference proxy: prompt policy, FinOps, observability and a central audit trail.

Maturity model

Most organisations sit at level 1 or 2

LevelArchitectureGovernance
1. ExperimentationGeneric agent + MCPNone
2. StandardisationGeneric + curated MCPApproved tools list
3. ProductionCustom + native toolingIdentity, RBAC, audit, evaluation
4. Enterprise-scaleCustom + orchestrationGovernance as code

Where MCP still wins

The mistake is MCP as the production integration layer

  • Prototyping: discover what an agent should do, fast.
  • Multi-vendor glue: tools with no shared platform.
  • Reference implementations: study MCP servers when building native tools.

For leaders

Agents are workloads. Treat them like it.

  • CIOs: one governed agent for your most painful workflow beats a hundred generic ones
  • CTOs: invest in agent-building capability, not the protocol ecosystem
  • Security leaders: minimise trust boundaries between agents and data

The takeaway

Not what agents can do, but what they reliably do, under governance, at scale.

Read the full article

Kevin Keller · kevinkeller.org