3 Days, 80 Commits, 5 Bounded Contexts
On a Friday morning in late May, I started designing an agentic runtime. Not a prototype — a production system. Five bounded contexts: gateway, dispatch, agents, knowledge graph, infrastructure. PostgreSQL task queue for deterministic orchestration. YAML-defined agents, each with single-purpose tools. An overnight pipeline that curates raw observations into a knowledge graph.
I had been running LLM routing experiments for five weeks before that — multi-model dispatch, cost tracking across projects, model selection logic. The routing layer worked. Building the orchestration around it felt like the obvious next step.
AI pair programming did most of the scaffolding. What used to take weeks took days. Eighty commits in three days. Five contexts built, boundary laws enforced, database migrations in place, CLI and Telegram adapters wired up. The system worked end-to-end: create a task, dispatch it to an agent, review the result, approve or reject.
Three days. It felt rational. AI made building cheap, so why use someone else's project?
Open Source Moved Faster
Eleven days later, a colleague dropped a link in our team chat. Databricks had just open-sourced Omnigent — a meta-harness that composes, governs, and shares AI agents. Sandboxing, policy enforcement, session collaboration, support for every major agent harness. Apache 2.0 license. Matei Zaharia called it a "harness of harnesses." It had 4,500 GitHub stars within two weeks.
I spent the weekend reading the docs. Omnigent covered composition, governance, and collaboration — three of my five contexts — backed by a team that would ship updates faster than I could fix bugs.
That same month: Microsoft Research released SkillOpt, a system that trains reusable agent skills without touching model weights — +23.5 average accuracy improvement across benchmarks. Google shipped ADK 2.0. PydanticAI hit 2.0 with full MCP and A2A support. The Agentic AI Foundation under Linux Foundation crossed 170 member organizations.
In May and June 2026 alone, open-source projects shipped more agentic infrastructure than most teams build in a year. Databricks open-sourced Omnigent, Microsoft Research released SkillOpt, Google shipped ADK 2.0, and PydanticAI reached 2.0 — each covering capabilities that take months to build in-house.
The runtime I built in three days? Commodity infrastructure. Covered by projects backed by teams of hundreds.
Build the 10% That Matters
I kept the routing layer. Killed the runtime.
The routing — model selection, cost tracking, multi-model dispatch — that was my actual differentiation. The orchestration, task queuing, agent sandboxing, governance? Infrastructure. Someone else will always build it better, maintain it longer, and test it harder than one developer ever could.
AI makes the trap worse. When you can scaffold a full system in three days, building from scratch feels rational. The prototype comes fast. The maintenance does not. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 — teams that built infrastructure instead of shipping product are the most likely casualties.
Before building an agentic runtime from scratch, spend one day researching what exists. As of mid-2026, projects like Omnigent, CrewAI, LangGraph, and Google ADK handle orchestration, governance, and multi-agent coordination out of the box. Build only the 10% that differentiates your product — the domain logic, the model routing, the part no open-source project will build for you.
The discipline is not in building. It is in checking what already exists, knowing what level of agent you actually need, and shipping on top of it.
The value was never in the architecture.
If you are evaluating how to ship AI agents without rebuilding the wheel, get in touch.