Field notes
Production AI patterns we ship every week. Stuck pilots, MCP hardening, eval suites, agent orchestration, cost control. Written by Alex Cinovoj, Founder and CTO of TechTide AI. No vendor pitches, no recycled blog posts.
22 of 22 field notes · 2026 archive
Copilot optimizes for keystrokes, Claude Code optimizes for tasks. A production comparison across agentic work, MCP tool access, review workflow, cost control, and governance.
Most AI ROI decks measure enthusiasm. Here is the baseline-first method we use on engagements: one workflow, one metric, cost line next to benefit line, reported monthly.
AI-generated prototypes are the cheapest product research available. The failure is promoting one to production without tests, types, auth, or an owner. The handoff checklist we use.
Cancelled agentic projects are rarely killed by the model. They are killed by unscoped ambition, no owner, no evals, no cost ceiling, and a use case that never needed an agent.
MCP is vertical, model to tools. A2A is horizontal, agent to agent. Most teams need the first and adopt the second too early. The decision rule, with failure modes.
A field-tested playbook: triage, harden, and ship one workflow to production. The sequence AI Production Systems follows once a decision justifies moving forward.
Six months after the first AI feature ships, the debt shows up as prompt sprawl, untyped tools, no eval baseline, and nobody able to say what changed. How we pay it down.
Model swaps stopped moving the needle. The gains now come from the harness around the model: durable state, checkpoints, retry policy, budget ceilings, and clean recovery.
The agent does not need to be right every time. It needs to know when it is out of depth. Four escalation triggers, plus the handoff payload that makes a human fast.
Leaderboards measure someone else's problem. Build a 50-task set from your own traffic, score it blind, price it per task, then keep the swap cheap.
Enterprise AI governance frameworks are unusable at 200 people. Here is the one-page version we install: data boundaries, approved models, logs, review gates, named owner.
Every long-context release restarts the RAG obituary. Retrieval still wins on cost, freshness, permissions, and citation. The five fixes that rescue a weak pipeline.
Filtering instructions out of untrusted text is a losing game. The defenses that hold are boring: least-privilege tools, typed arguments, allowlists, and approval on writes.
Swarms demo well and page you at 2am. Three orchestration patterns survive real traffic, and the choice is decided by failure isolation, not elegance.
Debugging an agent from a chat transcript is guesswork. Log the full run as a trace, keep cost and latency on the same timeline, and alert on tool-level failures.
Teams argue about whether the agent should be autonomous. Wrong unit. Set autonomy per action, using reversibility and blast radius, and earn each step up with evidence.
Nobody plans an AI budget overrun. They just never set a ceiling. Per-task budgets, tiered routing, context hygiene, caching, and one dashboard with a named owner.
Fourteen nodes in a state graph to answer refund questions. Size the stack to the problem: one loop, typed tools, a log, and only then a framework.
Every MCP server ships two products at once: the tools and the attack surface. Fourteen checks on auth, scoping, tenancy, argument validation, and audit trails.
No eval suite means every change is a vibe. Five days: harvest real tasks, write assertions, set the pass bar, wire CI, then gate deploys on it.
The prompt is the smallest part. Production quality comes from context assembly: retrieval scope, ordering, compaction, tool results, and a hard token budget.
MIT says 95%, Gartner says 40%, vendor decks say 88%. They measure different things. What each stat counts, and the five gaps that separate shipped from shelved.
Newsletter
© 2026 TechTide AI LLC. Columbus, Ohio.