AI & Agentic POV
The agentic stack, from someone shipping it
For Heads of AI, VPs of Engineering, agentic AI builders, and operators shipping production agents. Written from the Forward Deployed Engineer seat — what actually breaks in production, not what gets demoed at conferences. What's actually working in production, what's vendor theater, and where the next twelve months go.
Where Claude Code stops and senior engineers start
Generation went to zero. Reading, explaining, and judgment did not. The 2026 boundary between agent work and human work, from the Forward Deployed seat.
AI governance for teams that don't have a CISO yet
Most teams ship agents with shared API keys and no logging. Here is the minimum governance layer to build before you have a security org.
The agent failure mode taxonomy nobody publishes
The demo-friendly failure list is wrong. The four categories that actually kill production agents are stale data, missing context, hallucination, and conflict-with-truth.
Eval harnesses that catch calibration drift
Your agent's demo score is not lying. Your harness is. Here is how to design evals that surface the calibration drift quietly killing production agents.
The three-day MCP server design playbook
A senior operator's three-day sequence for shipping an MCP server that compounds instead of decaying into an API wrapper nobody wants to call twice.
Multi-agent is a coordination tax you may not need
Splitting a working single agent into a swarm usually buys an org-chart metaphor and pays in latency, failure surface, and debugging. When the tax is worth it, and when it is cosplay.
Your agent needs a budget, not just a prompt
Agents fail open on cost. No ceiling, no step limit, no breaker. Treat spend as a design constraint, not a dashboard you check after the invoice lands.
The model is not your bottleneck, your context is
Teams keep upgrading models to fix broken agents. The real constraint is what the model can see. Here is where the leverage actually sits.
MCP is a distribution play, not a protocol
The MCP spec is the least interesting part. The real fight is over who owns the tool surface agents reach for by default.
The eval you skip is the outage you schedule
Every agent incident traces to a failure mode someone called too rare to test. Here are the three evals that cover the majority of your real risk surface.
Most agent demos are just a for-loop with anxiety
A retry loop around one prompt is not an agent. State, tool choice, and evals are the line between a demo and production.
Custom MCP servers that compound vs ones that don't
Five practices that separate production-grade MCP servers from glorified internal scripts, written from the Forward Deployed seat.
Why "agentic" should mean operates, not responds
Production agents do not look like chat. They look like distributed systems. The operator pattern beats the chatbot pattern, and the CODN on getting this wrong is brutal.
The agent tax: hidden costs of bad MCP servers
Script-grade MCP servers feel free in the demo and bill you for years in production. Here is what they actually cost, and how to skip the retrofit.
AI governance is a logging problem before it's a policy problem
Enterprises write AI policy before they can observe what their agents do. You cannot govern what you do not log. Instrumentation comes first.
Context engineering is the job prompt engineering pretended to be
Prompt wording barely moves the needle in production. Context assembly does. Here is what actually breaks when you treat the window as a scarce resource.
Your agent doesn't need more tools. It needs fewer.
Tool sprawl is the silent killer of agent reliability. Cutting tools, not adding them, is what makes agents production-grade.
Evals, or it's vibes
Every production AI agent needs an eval harness. Without one, you're shipping vibes — and you'll only find out when the agent acts on bad data at scale, in production, repeatedly.
MCP is the protocol the agent conversation needed
Most 'AI agents' fail not because the model is wrong but because the agent has no way to read your data and act on your systems. MCP is the missing layer. Here's why it's the unlock.
Why Claude Code is the RevOps hire of 2026
RevOps is becoming a build job. Most teams can't hire fast enough at the new bar. Claude Code closes the gap — if you treat it like a junior engineer, not a chatbot.