Superintelligence for Startups: Habits That Compound Early

Superintelligence for startups is less about buying the most powerful model and more about building the habits that let any model work well inside your company: written knowledge, curated memory and repeatable evaluation. These habits cost little while the team is small, and each new model makes them more valuable because a stronger model can use more of what you recorded. Start with a plain operating document, a small memory an agent can read, and a set of real examples you can re-run.

Key takeaways

• Models change often, but your written context, memory and test examples stay yours and keep paying off.
• Documentation that a new hire could follow is also documentation an agent can follow.
• Memory should be small, owned and reviewed, not a dump of every chat and file.
• A short set of real examples with known good outcomes is the cheapest evaluation you can build.
• Skip heavy tooling for now. Habits first, infrastructure when the habits show where it is needed.

Why do habits compound while tools do not?

A tool you adopt this quarter may be replaced next year. A model you rely on today will almost certainly be superseded by a better one. What does not expire is the record of how your company works: who decides what, what good output looks like, which mistakes have already been made and corrected.

That record is what turns a general model into a useful colleague. A stronger model with no context still guesses about your customers, your tone and your limits. A stronger model with clear context uses its extra capability on your actual problems. This is why the value of early habits grows as models improve, instead of shrinking.

Startups have an advantage here. You have few legacy systems, few opinions that are only in someone's head, and a founder who can still say what the company is for. Writing it down now takes a fraction of the effort it will take after the team has grown and the knowledge has scattered. For a broader view of how this layer fits together, see Build a Company Superintelligence Layer That Improves With Models.

Which documentation habits are cheap to start?

You do not need a knowledge base project. You need a few short documents that are kept current and written for a reader who has no background.

• Company brief: what you sell, to whom, and what you will not do. One page.
• Voice and quality guide: examples of writing, replies or analysis you consider good, with a sentence on why.
• Role descriptions: for each recurring job (support triage, lead research, invoice checks), the inputs, the steps, the output and who approves.
• Decision log: a running list of choices you made and the reason, such as why you refuse a certain kind of request.
• Escalation rules: what must go to a human, and to which human.

The test for each document is simple: could a capable stranger do the job from this page? If yes, an agent can use it too. If no, the gap is the useful finding, because it shows where your process lives only in memory.

Keep these in one place with an owner and a visible last-reviewed date. A document that nobody maintains becomes a source of confident errors. Anthropic's engineering notes on [context engineering for agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) make a related point: what the model sees at the moment of work matters more than how much you can store, so short and accurate beats long and stale.

How should a startup structure agent memory?

Memory is the part most teams get wrong, usually by saving everything. A pile of old conversations gives an agent contradictory instructions and outdated facts. Good memory is chosen, labeled and reviewed.

Separate memory into three kinds:

• Stable facts: product details, customer segments, policies. These change rarely and should be edited by a named person.
• Working notes: current projects, open questions, what happened last week. These expire and should be pruned on a schedule.
• Learned corrections: when a human fixes the agent's output, record the correction and the reason. This is the most valuable memory because it captures your standards in concrete form.

For each item, note the source, the owner and when it was last confirmed. If you cannot say who is responsible for a fact, it should not be in memory the agent trusts.

A related choice is whether to teach the model your knowledge by retrieval or by changing the model itself. For most startups, retrieval from documents you control is the easier habit to keep, because you can read, edit and delete what the agent sees. AWS provides a useful [comparison of retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html) if you want the trade-offs in more detail. Memory kept as plain documents also moves with you if you switch models, which matters for keeping the gains in-house as models change.

How do you start evaluating without a data team?

Evaluation sounds like a specialist job, but the first version is a spreadsheet and some discipline. The aim is to answer one question every time something changes: is the agent better, worse or the same on the work we care about?

The step-by-step plan below fits in a few working sessions.

• Pick one role the agent will play, such as drafting replies to common customer questions.
• Collect real examples from past work, including easy cases, awkward cases and a few where the right answer is to escalate.
• For each example, write the expected outcome in plain words, not a perfect answer but the qualities a good answer must have.
• Run the agent on the set and have a human mark each result as acceptable, needs edits or unacceptable.
• Save the results with the date, the model used and the instructions in effect.
• Re-run the same set whenever you change instructions, memory, tools or the underlying model.
• Add a new example every time a real failure happens, so the set grows from your own mistakes.

This is the habit that compounds most. When a better model appears, you do not have to rely on announcements or intuition. You run your set and see what actually changed for your work. The [OpenAI evals guide](https://platform.openai.com/docs/guides/evals) and Anthropic's piece on [evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) describe more formal versions once you outgrow the spreadsheet.

Define what counts as success before you run anything. Our guide on success criteria for an AI agent pilot walks through how to make that observable.

What does this look like in a small team?

Here is an illustrative scenario, not a real company. A small software startup has one person handling customer questions and wants an agent to draft first replies.

In the first week, the founder writes a one-page brief, a short list of replies she is proud of, and a rule that refunds, security reports and legal questions always go to her. She assembles a set of past customer messages with notes on what a good reply needs to include.

The agent drafts replies, and the support lead edits them. Each time he changes something meaningful, he adds a line to the corrections file: what was wrong and what the rule is. Once a month, someone removes outdated notes from working memory and re-runs the example set.

Months later, the team tries a newer model. They do not debate it. They run the set, compare the marks and read the cases that changed. The new model handles awkward cases better but is too chatty for their voice, so they adjust the voice guide and run it again. The work took little time because the habits already existed, and the stronger model gained immediately from the corrections file.

What can wait until later?

Early habits do not require early complexity. Several things look important and can safely be postponed.

• A custom evaluation platform: a spreadsheet and a calendar reminder are enough until the example set becomes hard to manage.
• Fine-tuning a model: wait until you have a stable task, plenty of reviewed examples and a clear reason retrieval is not enough.
• Many agents at once: one well-documented role teaches you more than several half-defined ones.
• Broad tool access: give the agent only the tools and permissions the role needs, and extend them deliberately.
• A large governance program: a named owner, a written scope and a review rhythm are a sound start. See the superintelligence readiness checklist for a fuller sequence when you are ready.

The common thread is to keep what you build readable and reversible. If you can open it, explain it and change it, it will carry over to the next model.

How does Be Human's managed approach handle this?

Be Human designs, implements and operates AI agents with their identity, memory, tools and permissions under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. In managed operations, the documentation, memory review and recurring evaluations described here are part of the routine, alongside monitoring, permission review and controlled improvements. Your team keeps ownership of the content and decides what changes.

Frequently asked questions

Do we need superintelligence before these habits make sense? No. The habits are useful with today's models, and they are the reason you can adopt stronger ones quickly. Nobody can say when or how a more capable system will arrive, but a company with clear context, curated memory and a test set will be ready for it in either case.

Who should own documentation and memory in a startup? A named person per area, usually the founder or a team lead for the role in question. Ownership matters more than seniority: someone must be able to confirm a fact, retire an outdated note and approve a correction.

How many evaluation examples are enough to start? Enough to cover the common cases, the awkward ones and the situations where the agent should hand off to a human. A small set that you actually re-run is worth more than a large set that sits unused, and you can grow it each time a real failure shows up.

When should we ask for outside help? Ask when the habits exist but nobody has time to maintain them, or when the agent needs access to sensitive tools and systems. An implementation partner can set up the structure and run it, and what an AI implementation consultant does explains what to expect from that role.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.