Praetor Organization Operating System

Praetor treats an AI-run company as a governed organization, not a loose group of chat agents. The core objects below define how work moves from chairman intent to staffed execution, visible trace, review, durable memory, and closeout.

Mission lifecycle

Every mission has a current_stage and append-only MissionStageTransition records.

Stages:

The mission status still records operational state such as active, paused, ready_for_ceo, or archived. Stage records explain the business lifecycle.

Project portfolio

ProjectDefinition is the portfolio object above missions. A project represents a product, client initiative, research stream, or operating program. It has a name, tags, workspace path, stage, progress percentage, mission ids, budget/settings placeholders, routing notes, and token usage placeholders.

Projects can be created by the owner or by CEO. When a mission is created, the Project Router first tries to attach it to an existing active project by matching the mission title, summary, domains, requested outputs, repo context, project name, tags, and workspace path. If no project fits and the mission is not a trivial unclassified request, CEO creates a new project.

Project stages are intentionally simple for the dashboard:

Project progress is not state-weighted. It is the direct completion ratio requested by the owner: completed missions / total missions.

Agent employment contracts

When Praetor creates an agent, it also creates an AgentEmploymentContract. The contract records:

This gives the CEO and PM a stable way to tell an executor what it may do, when it must stop, and what "done" means.

Permission profiles

AgentPermissionProfile is the reusable safety envelope for agents. Defaults are:

Profiles intentionally separate agent role from permissions. A future LLM planner can recommend a role, while governance decides the permission profile.

Upward interaction requests

When an agent is blocked, it creates an InteractionRequest and sends it upward through the organization chart.

This is the default AI-manages-AI path. It prevents lateral ambiguity: the agent's manager owns the answer, escalation, or decision request.

Meeting ownership

Every AI meeting has exactly one owner of the meeting: the highest-level AI participating in that meeting.

The meeting owner controls the agenda, resolves disagreement, records tradeoffs, makes the operating decision, and assigns action items. If the decision crosses an approval boundary, the owner records the issue and escalates upward instead of pretending the meeting can authorize it.

Agent evaluation

Praetor evaluates agents by work behavior, not personality. Useful evidence includes completed tasks, review quality, unnecessary rework, clear questions, correct escalation, interaction answers, meeting decisions, and whether durable outcomes were written back into mission files, decisions, wiki, or memory.

The default routine is weekly Agent Development Review. A low review does not mean the agent should be removed; it means Praetor should produce an improvement plan. Each review records structured manager input and coworker/collaborator input so the evaluation is not only self-observation. The plan may add skills, update the employment contract, add review checkpoints, or reassign task types until the next review. CEO and manager agents can automatically apply internal improvements without owner approval, as long as they do not expand high-risk authority. High-risk permission expansion is converted into an owner approval request and linked back to the review instead of being silently applied. The owner can change the cadence in review policy.

Company evolution report

The CEO can prepare a weekly CompanyEvolutionReport from Agent Development Reviews. The report is owner-facing and describes how the AI company grew: which agents were improved, which skills were added, which guardrails were tightened, which risks were reduced, and what the company should observe next.

This is intentionally not a raw technical log. It gives Praetor a visible sense of company learning and maturity over time.

Company Playbook Review

Imported methods enter the Company Playbook as imported_requires_review. They are not assigned to new agents until reviewed. The UI and API can set a playbook entry to:

This keeps open-source method libraries useful without silently trusting external prompts or tool definitions.

Work trace

WorkTraceEvent is the chairman-readable event stream for AI-to-AI coordination. It records staffing, stage transitions, executor control requests, playbook reviews, and other operational events.

Raw agent messages and work-session turns remain available for audit. Work trace is the concise operating log that belongs in the main mission view.

Executive cadence

Praetor defaults to quiet formal cadence:

This matches Praetor's purpose: running serious company work through a CEO interface.