Notes from the agent frontier.

Handoffs, shared memory and parallel workers turn a simple trace into a graph. Here’s how we model it so a human can still read it at 2am.

Prompt changes now go through review like code. Sable 2.0 scores every pull request against your golden datasets and blocks regressions before they merge.

The best guardrails don’t just block bad outputs — they encode how your business wants an agent to behave when it’s unsure.

A look inside Sable’s ingest pipeline: columnar storage, adaptive sampling, and why we never drop an error span.

Our Series A, led by Northbeam with participation from existing investors, will fund self-hosted deployments, a European region and a lot more evals.

We analyzed 1.2 million failed agent runs. Almost all of them fall into seven patterns — and most are preventable with a single eval.