Notes from the Grove team
Engineering posts, design decisions, and field reports from teams running AI workflows in production.
-
When Software Starts Acting: Compliance in the Agentic AI Era
When software only computed, the code was the proof. Now that agents decide and act at runtime, you can't reconstruct what happened from source — so compliance has to be captured as it runs. What that takes, and how Grove builds it in.
-
Agents Shouldn't Grade Their Own Homework
GitHub built an impressive stack to validate non-deterministic agents — a prefix-tree graph, dominator analysis, a multimodal LLM. I needed the same thing for Grove's agents and found the hard part didn't exist. The reason is the whole argument for structured trajectories.
-
Harness as a service: ship the model, not the scaffolding
Most of your AI app isn't the model — it's the harness around it. Tool loops, streaming, sessions, audit, scheduling, integrations, interop. Grove ships that runtime as a service.
-
Reply first, finish later: respond_here for chat-shaped workflows
A flag that emits the caller-visible answer the moment the critical-path node completes — telemetry, ticketing, and audit keep running in the same DAG, on the same run id.
-
Build a self-improving essay writer in 50 lines
A generator/critic loop with RefineNode — write the accept criterion as a predicate, let Grove handle the iteration, the cap, and the rejection events.