Research / Open question

Can an organization learn without losing accountability?

We are exploring the organization itself as the unit of intelligence: people, AI agents, services, workflows, and shared knowledge operating with a common purpose and explicit authority.

From a signal to a useful response

An issue, message, or monitoring event enters the organization. The system preserves its source, relates it to current work, and identifies what needs attention. People and specialized agents consider responses. An authorized action is executed, and its outcome returns to shared knowledge.

Known routines can become deterministic playbooks. Consequential decisions still need accountable people and clear authority.

Current state

WebTree has concrete internal deployment and repository tools, plus early signal and attention artifacts. The broader federated organization remains an architecture and implementation direction. It is not a supported autonomous-company product.

MakeLocal and 0CAD are intended to operate as Project Cells with their own work and data boundaries. A Project Cell is an execution boundary; a Circle is a cooperation and governance boundary.

The company-as-brain analogy
MetaphorOperational meaning
ReceptorsEmail, messages, GitHub, monitoring, people, and agents
PerceptionNormalize inputs, resolve entities, and retain provenance
AttentionChoose what deserves limited time and reasoning
Working memoryA small shared set of active situations
World modelVersioned goals, commitments, dependencies, and evidence
Specialized reasoningDomain-specific interpretations and proposals
Executive decisionsResolve trade-offs and grant explicit authority
Playbooks and toolsExecute authorized routines with scoped permissions
Security / immune systemDetect and contain unsafe inputs or behavior
Global state / “hormones”Explicit priorities, separate from measured conditions
Memory and learningPreserve outcomes and improve future work

This is an architectural analogy, not a claim of consciousness.

What could go wrong?

Untrusted inputs can influence an agent. Shared memory can spread an error. Agreement among several agents is not independent proof. Missed or duplicated events can lead to inappropriate action, and automation can transfer hidden work to human reviewers.

Project data boundaries, authorization checks, replay and reconciliation, audit records, and ways to stop work all matter.

A useful next test

Choose one low-consequence recurring task. Compare a documented human workflow with an assisted one. Measure completion quality, reviewer effort, failures, and recovery from missed events. Keep a named decision owner.

The platform should earn its scope through useful project work. MakeLocal’s first customer should not depend on building this whole architecture.

Status and revision

This is a concept note, not a reported experiment result. Maintained by the WebTree project team, led by Max Levitskiy. Initial public synthesis: 5 September 2026. Future revisions should state what changed and why.

Bring evidence or a better test · All research questions