Writing from the team building the substrate — on state, memory, provenance, routing and validation, and what it takes to run agents you can audit.
Six products, one state graph, and a single idea — that an agent's work should be as reviewable as the code it writes. Here's what ships today.
Read more →Agents got dramatically better at doing the work. They got no better at explaining it — and that gap, not raw capability, is what keeps them out of production.
Read more →Five tools that sync with each other is five times the failure surface. One substrate with five views is a different architecture entirely.
Read more →Git records what changed and who typed it. For agent state that isn't enough — the commit itself has to carry intent, reasoning, confidence and authority.
Read more →Working memory and durable memory are different things. Confusing them is why your agent re-learns the same decision every Monday.
Read more →Ask why a function exists and the answer is in a Slack thread, a closed PR, or nobody's head. For agent-authored code, it needs to be in the repo.
Read more →Every chat UI makes you choose: keep this thread, or try the other approach. Branching a conversation the way you branch a codebase removes the choice.
Read more →Agents are non-deterministic, so a single passing run proves almost nothing. Validating them needs a harness built for distributions, not assertions.
Read more →Routing on token price optimises the wrong number. What you actually pay for is completed work — and three retries on a cheap model beats nothing.
Read more →"Context", "memory", "state", "agent" — four words doing twenty jobs each. A field that can't name its problems can't compare solutions to them.
Read more →