Documentation

What a genome is

A genome is a short list of plain-language instructions, called genes, that describe how one AI agent should work for one business. Examples: “Keep replies under 60 words.” “Never open an email with a greeting.” Before each model call, the genes that apply are written as a block of text and added to the agent’s system prompt.

A genome does not retrain or change the model. Its whole effect is text in the prompt, which means a model can still ignore it. The advantage of text is that a person can read it, edit it, roll it back and delete it.

A gene has a kind, a stratum (the layer it lives in), a strength between 0 and 1, a source and a list of references to where it came from. The full explanation is on Genes and strata. The wider picture is on the technology page.

The AGF format

The Agent Genome Format (AGF) is the file that carries a genome from one place to another. It is plain JSON. This site uses version 1.0. A genome pack on Nixtra is an AGF file plus a description.

Top-level fields

Genes and layers

Each gene has an id, a kind, text, strength, source, evidence_count, scope, overrides and provenance. It may also have a when condition and, for the role kind, a role_spec.

• Gene kinds: identity, approval, rule, role, decision, business, style, trait, tool_pref, fact_pointer.
• Strata a layer may use: org, team, customer, shared, user, learned. The kernel lives in the germline. The species stratum is never exported.
• Gene ids are 3 to 48 characters: lower-case letters, digits and underscore, starting with a letter.

Limits

• At most 200 genes in one file.
• At most 256 KB for the whole file.
• At most 400 characters of text in one gene.
• Provenance references must be in an allowed form. Execution ids, user ids and email addresses are never allowed.

The digest

integrity.sha256 is the SHA-256 of the document written as canonical JSON: keys sorted, no extra whitespace, with the integrity field removed. A reader recomputes it and refuses the file if the two differ. The digest catches accidental change and casual tampering. It does not prove who wrote the file, and it is not encryption.

Validate a file

The builder checks every rule above as you work and can import an existing file. The install page has a file checker and a list of common errors with their fixes. Both run in your browser. The file never leaves your device.

Built and tested G60 The format, its schema, the digest and the conformance files have code and tests in the IntelliNix reference implementation. Export and import in that platform are Built, flag off by default G61, G62 . See Portability.

How packs work

A pack is a genome that has been prepared for sharing. It is one folder in the catalog with a pack.json file. The file holds a name, a one-sentence summary, a description, a category, a version, a licence, the publisher, and the AGF genome itself.

• Kinds. A starter pack is written by Nixtra as an example for you to edit. A community pack is a submission that a person has reviewed. A partner pack comes from a named partner. At launch the catalog holds starter packs.
• Checks at build time. The catalog build stops if a pack’s genome fails validation or if its stored digest does not match the genome. A separate scan step checks pack text for injected instructions, personal data such as email addresses, secrets and unsupported claims.
• Regulated work. A pack for finance, health, law or similar work shows a notice: review it by a qualified person before use. This site gives no legal, tax or medical advice.
• Quality figures. A pack shows a quality figure only if one was measured, and then with its sample size, interval, method and source. Otherwise the pack page says “Not measured”. Starter packs have not been measured.
• Files. Each pack page offers the genome as a .agf.json download and shows two checksums: one for the downloaded file and the digest stored inside it.
• No ratings, reviews, download counts or prices. None of these exist on the site.

To use a pack, download it, read every gene, edit it for your business, and import it. Imported genes are checked again by the receiving platform, and learned genes arrive as proposals. The steps are on the install page and in the integration guides.

Where to go next