Field notes from shipping agents.
We publish what we learn — benchmarks, architectures, and methodology. No marketing posts. No takes you've already read.
- FrameworksMay 3, 2026
Building agents with LangGraph — a field guide
The most-used production agent framework in the field. Mental model, primitives, durable execution, and the named production users (LinkedIn, Klarna, Replit, Uber, Elastic).
- FrameworksMay 3, 2026
Building agents with the OpenAI Agents SDK — a field guide
Five primitives, a handoff model that just clicks, and tracing wired in by default. Mental model and where it earns its weight on OpenAI's stack.
- FrameworksMay 3, 2026
Building agents with CrewAI — a field guide
Role-based multi-agent. Hire a team — researcher, analyst, writer — and let the framework run the collaboration. The Crews-vs-Flows decision and where it earns its weight.
- FrameworksMay 3, 2026
Building agents with Microsoft AutoGen — a field guide
A complete actor-model rewrite of the framework that put multi-agent conversation on the map. The mental model, the layers, and the honest answer to whether you should still use it.
- FrameworksMay 3, 2026
Building agents with the Claude Agent SDK — a field guide
Anthropic's SDK is the same harness powering Claude Code, exposed as a library. The mental model, the primitives, the hook layer, and where it sits next to Google ADK.
- FrameworksMay 3, 2026
Building multi-agent systems with Google ADK — a field guide
ADK is the framework you reach for when chained prompts stop being enough. The mental model that makes it click, the four agent types you actually use, and where it stops.
- MethodologyMay 3, 2026
Evals as a deliverable: how we ship eval harnesses with every agent
Golden sets, pass/fail thresholds, regression tracking, and a public methodology page that buyers actually read.
- ArchitectureMay 3, 2026
Why your agent belongs on the phone, not the cloud
The privacy, latency, and cost case for on-device-first agentic architectures — and where cloud still belongs.
- MCPMay 3, 2026
Shipping an MCP server for a cooking workflow — what we learned
A walkthrough of designing tool boundaries, handling streaming results, and publishing an open-source MCP server.