One agent building software has two problems, and neither is the model. It cannot specialize, and it cannot remember past its session. The fix is a team of specialized agents in visible tmux panes, a builder, an adversarial auditor, and a cicd agent, coordinating over a message bus and grounded in a shared YAML story versioned in the repo. The bus is for coordination, the files for grounding.
Building a non-deterministic system is iterative tuning, and the knowledge it produces is fragile, it lives in your head and the chat scrollback, both gone the moment the session ends or the agent crashes. Evaluation-Driven Development makes that context durable by writing it to files, one YAML story per investigation, versioned in the repo, with an append-only run log the next agent reads.
Model infrastructure is unreliable, and not only at the edges. Calls time out, come back empty or truncated, or arrive well-formed and wrong. A pipeline built on model calls needs three things wrapped around every one, retry for the transient failures, an audit that rejects a response that parses but is wrong, and a trace so nothing dies in silence. A model call wrapped in all three is a durable task.
A large language model is brilliant and knows nothing about your business. Files are how the platform closes that gap. Every document, image, audio and video a customer uploads is turned into text and then embeddings, all of it landing in one searchable meaning-space held in the same Postgres as everything else. Everything the AI writes is reasoned from that material and carries the id of the source file and chunk it was grounded in, so a claim can point back at the exact paragraph that justified it.
If the cost of serving a customer changes with every page they generate, what goes on the invoice? Not a seat, which consumes nothing, and not an invented currency bought in blocks, which is the standard escape and one I did not take. The bill is in dollars, the meter is in tokens at a fixed rate per million, and both are on the same screen. A monthly cap the customer sets is what makes showing them the real number safe.
Giant Context writes your knowledge base, the help articles a customer reads to answer their own question, out of your own material. Writing them is one job. Keeping them true as your product changes is the one that never stops. Mind reads your articles back against what your business looks like now, finds the button that moved and the feature that was renamed, and proposes a specific fix to a specific article for you to approve. It names the article, names the problem, and waits.
A normal CRM activity log is a set of form fields, type, subject, outcome, due date, written for a salesperson to scan. I threw that out and reshaped the log into a natural-language timeline, one plain sentence per event, with the exact references in a structured slot on the same row. Mind reads a contact's timeline oldest to newest and works out what should happen next. If something other than a person is going to read your data most often, that is who the schema is for.
Giant Context sends your marketing email with no campaigns and no segments. Every other email tool is built on both. You write, in one sentence, when an email should go out, and Mind works out who gets it and when from your own contact activity. Instead of campaigns competing for the same inbox, each person gets one ordered stream, one message at a time.
The same files that build your marketing site do a second job you never ask for. They make it findable. You should not have to know what a sitemap is, or write a meta description, or think about structured data. Every site ships with a sitemap, canonical and language alternates, meta descriptions written from the page's own content, and FAQ structured data, in every language it renders. You never shopped for it. It came with the site.
A model handed nothing but a request invents. A warm testimonial no customer gave, a feature the product does not have. Giant Context builds your marketing site from your own files instead. Every concrete claim the model proposes has to anchor to your file content above a similarity threshold, and anything that does not is dropped before it reaches the page. The result is a site you can stand behind.