Agentic Development Enablement

Trained 400+ engineers across 50 teams and deployed over $500K in AI tooling, taking R&D from zero AI-assisted output to over two million lines of agent-produced material in two quarters.

The problem

AI adoption across R&D was fragmented. Some teams had dabbled with one-shot prompting, most had not touched agentic tooling at all. There was no shared platform, no training infrastructure, and no organizational standard for how engineers should work with AI. Leadership could not distinguish signal from noise, and every team that did experiment was paying the same onboarding tax in parallel.

The approach

The program followed a procurement-first, training-second model: tools in hands before theory on slides. A broader shortlist of coding assistants and agent platforms was evaluated and narrowed to two selections: the Cognition suite (Windsurf, Devin CLI, Devin, DeepWiki) and Anthropic Claude Code. Rejected options did not meet the enterprise-readiness bar required for formulaic rollout, and lacked the agentic depth the program was targeting.

Each procurement started with a proof-of-concept period. A smaller group evaluated the tool, developed best practices, and became the leaders and mentors who then trained the rest of the organization. This PoC-to-mentor cascade meant early adopters carried the credibility and context needed to bring skeptics along.

Training ran in every format that met engineers where they learned: office hours, formal sessions, innovation-hour showcases, hands-on labs, and vendor-led workshops with the tool creators themselves. Smaller working groups and async cohorts handled the long tail. Detailed written documentation served as the reference backbone.

Three audience archetypes required different approaches. Skeptics, rightly cautious that not all AI products deliver on their promise, needed help finding where AI genuinely fit their day-to-day. Senior engineers convinced manual work was faster were brought into the PoC model. Over-eager junior engineers were trained in structured prompt-and-context engineering, shifting focus from prompt input to the context provided to the agent and from outputs to outcomes delivered incrementally and algorithmically.

Alongside training, the program built an internal marketplace of reusable skills, hooks, agents, and workflows, and promoted a contribution culture where shared agents could help teams across the organization. Each entry in the marketplace carried metadata that made it discoverable, composable, and safe to invoke across squads:

# A skill from the internal ai-tools library, callable by any squad's agent.
name: compliance-scan
description: Triage product-compliance findings in a PR diff.
owner: architecture-services
tools: [read, grep, bash]
loop: observe-plan-verify

Reliability was defined as repeatability under scientific-method rigor. The program adopted a universal problem-solving loop as the foundational algorithm: Observe, Think, Plan, Build, Execute, Verify, Learn. The loop provides consistent state and context management, forces verification before declaring a result, and prevents hallucinations from compounding. A workflow clears the bar when the loop runs end-to-end, repeatedly, measurably, and without error.

The outcome

Over two quarters, the program trained 400+ engineers across 50 teams and deployed over $500K in AI tooling. The organization went from zero AI-assisted output to averaging over two million lines of agent-produced material spanning software development, documentation, data analytics, and other use cases. That represented a 10× multiplier on individual contributor workflows.

One team responsible for implementing governance guardrails across agentic integrations was bootstrapped with the same tools it was governing. Within months, the team was using agents end-to-end to ship the governance controls protecting data in source systems: research, analysis, planning, development, testing, verification, and delivery. The meta-loop became the program’s clearest proof point.

The assumption going in was that engineers wanted AI to replace their work. The reality was different. What they needed was relief from the mundane: the repetitive tasks that drained capacity without engaging judgment. Once AI took on drafting, note-taking, and boilerplate, engineers redirected that time toward data-driven decisions, creative problem-solving, and the innovative work that had been squeezed out by volume. The value was never in what AI could do on its own. It was in what it freed engineers to do instead.

What I’d change

Context engineering training should have preceded tool distribution. Teaching engineers how to scientifically structure their workflows and prompts, defining guardrails for the AI and a clear path toward testable outcomes, would have raised adoption quality from the start. Without that foundation, early usage defaulted to unstructured prompting, and the program spent time correcting habits that structured training would have prevented.