Agentic AI built for your operations.

Common consulting engagements

From discovery to deployment

How we approach agent architecture in engagements

The patterns behind these consulting engagements are documented openly on the blog. For cost discipline, start with pre-agentic data fetching and the 19–62% token cut and the prompt-caching playbook that drops repeat-call cost 47%. For model tiering before you touch routing code, the visual note stop paying frontier prices for classification is the map I hand to teams on day one. For orchestration, read why I'm replacing supervisor patterns with handoffs in multi-agent systems. And for the registry hygiene that drives tool-selection accuracy in production, tool descriptions are prompts — fix the registry, not the agent. Before any of that, the first audit on every engagement is the tool contract itself: your agents aren't broken, your tools are runs the three questions (atomic, honest on failure, typed) I check before touching a prompt. For agent memory architecture — the layer most production agents are silently missing — the three paradigms of LLM memory (implicit, explicit, agentic) is the map we use in every engagement. For the whole picture in one place, the anatomy of an AI agent is the four-box diagram (memory, tools, the loop, guardrails) I draw at the start of every engagement.

Every engagement also ships with the agent observability stack we deliver to every client and eval datasets that go beyond the happy path. For a one-glance carousel summary of the tool-design rules, see the visual note on fixing wrong-tool calls.

Before committing to an agent architecture at all, read AI agent vs agentic AI and what the distinction actually means when you ship one. That post is the framing I use to scope every engagement and includes the three-question test for which architecture your project actually needs. Once the architecture is settled, the next decision is execution shape: code agents vs skill agents and when to pick which walks through the keyboard-vs-toolbox tradeoff and the hybrid pattern most production systems become. For the implementation layer that lives inside both shapes, tool registry design for agentic AI is the audit recipe I run on every client codebase before changing anything else. For a worked example of all these patterns stitched together, Recruiting Atelier is the runnable reference I built: ReAct loop, supervisor with planning, tool registry, MCP integration, RAG, guardrails, observability — every primitive wired up in plain code you can read in an evening. For a planning-layer architecture at enterprise depth (12 stages, 17 guardrails, task harnessing, IDE prompts), see the Planner Agent reference spec.

On the model-routing and governance moves shipped in the May 2026 window: for the routing-layer question after Google\'s Gemini 3.5 Flash launch, Gemini 3.5 Flash vs Sonnet 4.6 and the cost-per-completed-task maths is the migration recipe I am running at clients this week. For enterprise MCP governance after Databricks shipped Unity AI Gateway, what changes for enterprise agents and the migration I would run for a Databricks shop covers the four-primitive build-vs-buy line. And on the protocol itself, the May 21 MCP spec release candidate takes the protocol stateless which is the architecture shift I am factoring into every new MCP build. Security moved too: after a poisoned VS Code extension harvested Claude Code credentials, hardening the agent supply chain (extensions, skills, MCP configs, and keys) is now the first review on any production engagement.

On the June 2026 window for billing, governance, and grounded retrieval: for the same-day Claude plumbing cutover (Agent SDK credit split plus Opus 4 / Sonnet 4 API retirement), the June 15 billing checklist I run before anything breaks at 2 a.m. covers auth paths, credit claims, and model ID greps. For Cursor Auto-review and pre-push /review as an autonomy dial (not a security boundary), governing agent autonomy in 2026 is the adoption path I am rolling out on client repos this week. For frontier routing after Fable 5 shipped and suspended in the same week, Claude Fable 5 for agent builders covers retention, safeguard fallbacks, and fallback IDs. And for Google's Agentic RAG preview with a Sufficient Context Agent, when iterative retrieval beats retrieve-then-pray is the decision table I use before re-architecting a knowledge base. For content QA pipelines (GEO + SEO eval before publish), the visual note agentic AI content quality: 5 agents, one pipeline is the architecture I run on pages before they go live.

For the full landscape, browse the topic pillars: Agentic AI, multi-agent systems, Model Context Protocol, AI observability, and enterprise AI automation.