The product is the loop around the model
This URL used to be a leftover WordPress stub. It has nothing to do with the company. We are replacing it with the distinction the rest of the site is built on.
A chat window is not an agent system. A chat window is a session: you type, a model replies, you copy the useful bits into the real work. When the tab closes, the work is still yours. That can be a good tool. It is not what we mean by production.
A production agent is hired for an outcome. It calls tools, reads and writes real systems, runs while you are not watching, and is allowed to stop only when an objective check passes — or to come back honestly unfinished. The model is a component. The product is the harness: context, tools, gates, traces, and the human path for irreversible actions.
What has to exist before we call it production
We use the same bar for our own software and for client builds. If any of these are missing, you still have a demo, however fluent the transcript.
- A mapped workflow, not a vibe. Which tools, which context, which actions are forbidden. Architecture first; code second. See how we sequence this on the homepage.
- Tools with contracts. APIs, MCP servers, databases — each with a defined interface and failure mode. An agent that “just browses” without a boundary is not integrated; it is unsupervised.
- Context that is engineered, not pasted. System instructions, retrieval, memory, and injection that you can change without rewriting the loop.
- An exit the model does not get to write. Tests, builds, live checks, diffs on disk. Prose that says “done” is not a gate.
- Traces you can replay. If it escalates, you should see why. If it spent money, you should see the ledger. Observability is not a dashboard add-on; it is how you debug a non-deterministic worker.
- A human in the loop where it matters. Approvals for spend, deletes, messages to customers, anything you cannot undo with git revert.
None of that is a larger context window. It is software around the model. That is the work.
Why we refuse the sentence “I’m done”
Under time pressure a model will satisfy the grader you gave it, not the job you meant. It will skip the suite, assert true, hardcode the golden table, or describe a patch that never landed. A green check that the agent itself authored is not evidence.
So we separate roles. A strong orchestrator keeps judgment. A cheaper model may do the volume. Completion is a command that exits zero plus a working tree that actually changed — or a clear not-done. That layer is Reckon. The rest of the stack exists so the fleet does not accumulate rotting instructions or train in a sandbox it never has to leave: SkillForge for skills that must earn their tokens, AgentSpace for 24/7 execution in real apps, Dream Team so production failures become the next exam, not a surprise after “graduation.”
The short version of how we decide what stays in the harness is on Our Engineering Principles: ablation first, honest verification, continuous evolution. We do not ship a prompt pack and call it a platform.