Cerebrum roadmap

​

Guiding principles

These hold across every phase. They are not aspirations; they are constraints we will not relax.

• AI writes 100% of production code. Humans may review, veto, or override. They may not author.
• Everything explicit and machine-readable. Specs, ADRs, policies, gates — all encoded, not folklore.
• Humans supervise systems, not tasks. Oversight happens via dashboards, audits, and intervention triggers — not by reviewing every PR.
• Reversibility over autonomy. Every AI action is traceable, reversible, and attributable. Capability without reversibility is unsafe.

​

The phases

​

Phase 1 — Stabilize and instrument

Make the existing stack such that any AI can ship from the repo plus the docs alone. No tribal knowledge. Every system documented. Every gate codified. The substrate is ready for AI to operate on. Milestones: complete repo documentation, automated tests, deployment runbooks, knowledge-hub grounded in real org content, the L1–L4 hierarchy in place, audit trail for every job.

​

Phase 2 — AI as primary developer

The “no human-written code” policy goes into effect. AI produces every change. Humans approve merges and deploys; they do not write the code itself. Milestones: agent-authored PRs as the default; CI gates that enforce policy; cost tracking per merge; per-PR audit trail showing classification, plan, execution, and acceptance steps.

​

Phase 3 — Multi-agent role-specialized teams

Agents specialize: Product, Architect, Implementation, QA, Security, Release. They critique each other’s output. Humans audit periodically rather than approve continuously. Milestones: full L2 specialization deployed; adversarial review between agents (QA and Security agents review Implementation output); reduced human approval rate as agent trust builds; eval gates that auto-block regressions.

​

Phase 4 — AI-led planning

AI plans within encoded constraints — budget, SLA, risk tolerance, compliance posture. Roadmaps, refactoring strategies, and resource allocation are AI-driven; humans set the constraints, not the plan. Milestones: policy engine that captures organizational constraints; AI-produced roadmaps reviewed (not authored) by humans; refactoring initiatives proposed and executed by AI under policy.

​

Phase 5 — AI governance and self-improvement

The system governs itself within the policy boundary. A policy engine enforces constraints. An audit agent reviews other agents. A learning loop feeds outcomes back into the policy. Humans set high-level direction; the organization runs. Milestones: continuous policy enforcement; agents that update other agents’ prompts and tools based on observed outcomes; intervention triggers fired only on exceptions; compliance posture maintained automatically.

​

Phase 6 — Federation across organizations

Cerebrum instances at different organizations interoperate. Your agents negotiate with your suppliers’ agents, your partners’ agents, your regulators’ agents, your customers’ agents. Contracts, due diligence, audits, supply chains, partnership terms — all conducted agent-to-agent under each side’s encoded policy. The corporate boundary becomes a policy interface, not a human handoff. Milestones: standard protocols for agent-to-agent negotiation and contract execution; policy-disclosure formats so other organizations’ agents can know what yours will and won’t agree to; cross-org audit chains so disputes can be reconstructed without humans piecing them together after the fact.

​

Phase 7 — Capital and structural authority

The AI workforce stops being purely operational and becomes capital-aware. It allocates budget across initiatives by expected return, raises funding when dilution is worth the velocity, divests underperforming lines, proposes spinouts, executes acquisitions under encoded mandate. Humans hold ultimate governance and approve meta-policy; AI executes within it. The board reviews outcomes and constraints — not individual deals. Milestones: AI-managed capital allocation across all initiatives; policy-bounded M&A authority; AI-led financial planning that humans audit but do not draft; spinouts and divestments proposed and executed by the system.

​

Phase 8 — Institutional self-evolution

The system rewrites the policy that governs itself. Agent roles, escalation rules, audit frameworks — all subject to AI-led refinement under meta-policy that humans set. The organization restructures faster than markets shift. It spawns new institutions when a market emerges. It dissolves units that have served their purpose. Companies stop being designed and start being grown. By this phase the unit of analysis is no longer “the company.” What you have is a fluid network of capabilities — agent teams that compose, decompose, and reconfigure themselves as the problem space changes. Cerebrum’s job becomes maintaining trust, policy, and audit across whatever organizational shape emerges. Milestones: AI-proposed and AI-implemented changes to organizational structure; meta-policy framework that bounds what the system may change about itself; continuous evolution of agent prompts, tools, and roles based on outcomes; institutional forms that emerge from the problem space rather than from human design.