A-Genetic Engineering

A-Genetic Engineering is the name for how The One writes a multi-agent team down: in full, as files in a Definition Space that you can reuse, edit and pass on. Prolifera Engineering describes how the system grows while it works; A-Genetic Engineering describes what it grows from. The A stands for Agent: Agent-Genetic, abbreviated.

Two half-answers

The previous generation of Agent products each held half of the answer:

• Single-body Agents like Claude Code and Codex use Skills to record verified experience and procedures. That is experience, but with no structure to carry it. In each new conversation a Skill lands in a different context, executed by a different implicit role, so the experience drifts.
• Their multi-agent stories are ephemeral sub-agents: spawned for chores, discarded when done. That is structure, but with no heredity. However sound the division of labour, it dissolves when the run ends, and the next run starts from zero.

Experience without structure drifts; structure without heredity dissolves.

The genotype: what a Space writes down

In The One, a team is written down in full in a Definition Space, and you find each part of it under Teams and tools, where each working group's card opens its template:

• Working-group templates (workflows/*.yaml, the Graphs): nodes and edges. Each node binds a role and can add instructions for that node alone; each edge's text is the handoff instruction: when to hand off, what to pass, what the other side should produce. This is the team's body plan.
• Roles (agents/*.yaml, the Agent presets): each member's instructions, its model preset or the default model, reasoning, built-in tools and MCP servers, Skills, prompts and Plugins, defined independently of every other member. These are the cell types.
• Skills and Prompts: written-down ways of working and shared instructions, mounted on whichever roles need them.
• Plugin references: a role can name the Plugins it uses. The Space keeps only the reference; the Plugin itself is installed on each computer, outside every Space.
• Memory: lessons and the reasons behind the Space's design, written by Agents and by the nightly review. Each turn, an Agent that holds the memory tools sees up to 20 entries, the ones most relevant to its latest messages. See Memory and recall.

Taken separately, the previous generation had prototypes of all of these. The fusion is what matters. When an experiential Skill lands in a fixed position (a role with explicit instructions, explicit tools, an explicit place in the team and explicit handoff edges) rather than in whatever conversation happens to load it, experience stops drifting. When the structure and every member's full configuration are written to files rather than scattered across one run's memory, structure becomes heritable. A mature, reusable team is the fusion of the two, and this complete record is its gene.

The phenotype: a working group on a matter

Genotype and phenotype come apart. The template is the genotype. When the main assistant gives the work of a matter to a template, The One starts a new run of it: a working group, the phenotype, at work on that matter's board. One template can be at work for several matters at once; on Teams and tools, its working group's card names them, as Working for 2: followed by the matters' titles.

Expression can continue inside a group: with create_subgraph an Agent runs another template as a sub-Graph of its group, and can open that part of the work as a sub-matter with its own board. Groups sleep, pause and end; the gene remains, and the next expression reproduces the structure and configuration verbatim. (Faithful means the configuration is reproduced, not the behaviour: model output is still stochastic, and one gene does not guarantee two identical runs.)

The environment can change without touching the gene. The main assistant can give one group's run other models for that piece of work alone. Binding a model preset to a role or a template changes every later run, so it is kept for when that is the point.

The main assistant is the genetic engineer

None of this has to land on you. The main assistant of your Space decides which gene to express and when one should change:

• Selection. For each matter it picks the template that fits from the Space's templates, which are listed in its context with what each is for, and prefers one that has done this kind of work well.
• Authoring. When no template fits and the work will come back, it has a new one written: it gives the authoring to the Definitions studio, the atlas working group, or edits the definitions itself with the definition-authoring Skill.
• Observation. It holds each delivery against what you asked for, answers a group's questions or leaves them to you, and hears when a group stops. The verdict on a delivery stays yours: Accept, Rework or Give up.
• Directed evolution. When you ask the Space to get better at its own work, it opens the evolution matter (below).

The editing bench underneath is the same for every Agent. It reads the whole Space at its exact revision, composes a batch of changes, can dry-run the batch and validate it against the schemas, the references between definitions, tool and MCP bindings and Plugin availability, and then commits it as one change. Every edit an Agent makes is recorded in the Action log and can be undone. The editors under Teams and tools and the Atlas change the same files.

Directed evolution

Gene editing driven by evidence is directed evolution, and The One runs two loops of it in each Space:

• The evolution matter. You ask the main assistant to improve the Space. It opens a matter called Evolution and gives it to the lab working group: Director, Patcher and Verifier. Each round starts from a census of real failures, tests at most one hypothesis, with its bar written down before the first reading, changes the smallest layer that failed, and ships only what clears that bar, with the old version frozen so it can be taken back. A weekly census keeps the matter going and skips weeks with nothing new.
• The nightly review. Every night at 03:30, when real use shows new friction, a Reviewer turns what repeats into a lesson in memory or a better Skill, at most three changes a night. Each change is measured against the friction it targets, before and after; the Reviewer undoes or improves a change that made things worse or did nothing, and you see every change in the morning report, where you can undo it.

Self-Evolution describes the two loops as an idea; Evolution: the lab and the nightly review shows how to start and follow them in the app.

Heredity

Genes matter because they are passed on. A Space can be copied, exported, imported and shared (.zerospace), and its definitions travel with it:

• Memory travels when you choose: export includes agent memory unless you untick it, and sharing leaves it out unless you tick Include agent memory.
• Plugins travel as references; the computer that receives the Space installs them separately.
• Secrets never travel.
• Models stay on each computer too: roles name model presets, which live in the app, so a Space from elsewhere may need its roles pointed at your own presets.

The official Space is the read-only baseline genome: in it, what the main assistant and the evolution lab find stays a proposal. Copy it into a Space of your own to breed variants; later updates to the official Space do not reach your copy by themselves. See Definition Spaces.

Boundaries

Like Prolifera, the analogy is borrowed from biology as plain vocabulary for the mechanisms, not a claim that Agents are alive. Some boundaries worth stating directly:

• The host does not edit genes by itself. Apart from adding built-in definitions a Space is missing (it never overwrites a file), every definition change is an Agent's decision or yours. The nightly review runs in the background, but what it changes is the Reviewer's decision, recorded, measured and undoable like any other edit.
• Genes do not guarantee performance. They fix structure, configuration and where experience lands, removing drift between expressions; the quality of any single run still depends on the model.
• An edit is not a validation. Whether a change helps is settled by readings: the lab's bar written down beforehand, the nightly review's before-and-after counts, or running the work again and looking. It is never assumed.

The concept map

Start from Single Agent: From Single Agent to Multi-Agent explains why complex work needs several Agents with clean roles; this page explains how that collaboration is written down as heritable genes; Prolifera Engineering explains how genes are expressed and grow at runtime; Self-Evolution explains how they are improved, deliberately, from failure.