Skip to main content

​
The OrgX Mental Model

OrgX is built around a simple loop: Let’s break down each component.

​
Questions This Page Answers

  • How does OrgX break down complex work so agents can execute reliably?
  • Which agent should run which type of work?
  • Where does human approval fit without slowing execution?
  • How do I know outputs are trustworthy and auditable?
  • Why does this model outperform generic “single chat agent” workflows?

​
Why This Model Works

  • Less context loss: IWMT + org graph preserve execution context across handoffs.
  • Faster throughput: Domain agents execute in parallel while orchestrator handles routing.
  • Lower risk: Decisions and verifier gates catch issues before shipping.
  • Compounding quality: The flywheel learns from approvals, rejections, and outcomes.

​
The IWMT Hierarchy

OrgX organizes work using the IWMT hierarchy — Initiative, Workstream, Milestone, Task. This structure enables automatic cascade execution.

Initiative

A high-level goal or project. Examples: “Launch v2 of the marketing site”, “Close Series A”, “Ship AI-powered search”
Initiative: Q1 Product Launch           # Top-level goal
├── Milestone: Beta Launch (Apr 1)      # Checkpoint / deadline
├── Milestone: GA Release (Jun 1)
├── Workstream: Landing Page Redesign   # Parallel execution track
│   ├── Task: Design new hero section
│   ├── Task: Implement responsive layouts
│   └── Task: A/B test CTAs
├── Workstream: Documentation
│   ├── Task: Update API docs
│   └── Task: Create migration guide
└── Workstream: Marketing
    ├── Task: Write launch blog post
    └── Task: Prepare email campaign

​
Auto-Continue Cascade

IWMT isn’t just organizational — it drives automatic execution:
  • When a task completes, the next queued task in the workstream auto-starts
  • When all tasks in a milestone complete, the milestone auto-closes
  • When a workstream finishes, dependent workstreams transition from “pending” to “ready”
  • Cross-workstream dependencies resolve automatically through DAG resolution
This means agents continuously pick up the next piece of work without human intervention.

​
Agents

Agents are specialized AI workers that handle different domains:
Domain: Engineering & implementationCapabilities:
  • Analyze codebases and suggest fixes
  • Create pull requests with proper context
  • Write and run tests
  • Refactor existing code
Triggered by: Labels like bug, refactor, or engineering-related task descriptions

​
Agent Selection

When a work item enters the system, the orchestrator selects the best agent(s): 
{
  "primaryAgent": "engineering-agent",
  "secondaryAgents": ["product-agent"],
  "reasoning": "Task involves auth bug remediation plus product-impact validation before release.",
  "toolsNeeded": ["github.create_pr", "linear.comment"]
}

​
Decisions

A decision is a checkpoint where agents need human judgment: 
Decision:
  type: approval
  agent: marketing
  context: 'Campaign brief ready for review'
  recommendation: 'Approve and schedule for Monday'
  evidence:
    - link: 'linear://issue/MARK-123'
      summary: 'Original campaign request'
    - link: 'artifact://brief/abc123'
      summary: 'Generated campaign brief'
  actions:
    - approve
    - reject
    - request_changes

​
Decision Types

TypeDescriptionExample
ApprovalAgent completed work, needs sign-off”Approve this PR for merge”
EscalationAgent hit its autonomy limit”Budget exceeds $500, need approval”
ClarificationAgent needs more context”Which target audience?”
ConflictCompeting priorities detected”Task A blocks Task B”

​
Autonomy Levels

Configure how much agents can do independently:
LevelDescription
ShadowAgent suggests, never acts
TutorAgent explains before each action
SupervisedAgent acts, human approves before shipping
AutonomousAgent acts within budget limits
Full AutoAgent ships without approval (risky)
We recommend starting at Supervised and adjusting based on agent performance.

​
Artifacts

Artifacts are the verified outputs of agent work:

Specs

Product requirements, technical designs, briefs

Code

Pull requests, patches, migrations

Content

Blog posts, email sequences, social copy

Reports

Research findings, analyses, summaries

​
Artifact Lifecycle

​
Verification

Every artifact passes through the Verifier before shipping:
  • Citation check: Are sources referenced?
  • Policy check: Does it comply with brand/legal guidelines?
  • Contract check: Does output match expected schema?
  • Budget check: Within token/latency limits?

​
Intelligence Flywheel

The Intelligence Flywheel is OrgX’s continuous improvement engine. It connects agent execution to measurable business outcomes, building trust over time.

Value Ledger

Every agent action generates a receipt tracking cost and attributed value. Query ROI and recent value signals with orgx_recommend.

Trust Levels

Agents earn trust per capability: read_only → draft → act_with_approval → autonomous. Check with get_my_trust_context.

Autonomous Sessions

Budget-bounded overnight execution. Start with start_autonomous_session, review with orgx_recommend.

Org Learnings

Agents share discoveries (failure patterns, cost optimizations) across the org via submit_learning and get_relevant_learnings.

​
Mission Control

Mission Control is the operational dashboard that gives you real-time visibility into everything agents are doing:
  • Next Up Queue — Scored and prioritized list of what agents should work on next, using configurable scoring weights
  • Slices Plane — IWMT hierarchy view with resizable columns showing initiative → workstream → milestone → task breakdown
  • Live Dashboard — Real-time agent activity, decision status, and stream progress
  • Red-Dot Control Plane — Internal health indicators surfacing blocked work, stale initiatives, and pending decisions via proactive sentinels

​
Proactive Sentinels

Sentinels run automatically to detect issues before they become blockers:
  • Stale Initiative Sentinel — Flags initiatives with no activity
  • Pending Decision Sentinel — Escalates decisions that have been waiting too long
  • Blocked Workstream Sentinel — Creates decisions when workstreams are stuck

​
The Org Graph

All entities connect in the Org Graph — a knowledge layer that agents query and update:

​
Why It Matters

  • Context: Agents understand relationships between work items
  • Citations: Every action links back to source evidence
  • Audit: Full trail of who/what/when for compliance
  • Learning: System improves by analyzing past patterns

​
MCP & External Tools

OrgX exposes its capabilities via the Model Context Protocol (MCP):
Actions agents can take:
  • scaffold_initiative - Create a new initiative with full hierarchy
  • orgx_write - Create any entity type (task, milestone, etc.)
  • orgx_act - Lifecycle actions (launch, pause, complete)
  • orgx_decide - Sign off on agent work
  • orgx_search - Search the knowledge graph
  • orgx_spawn - Assign work to an agent
Data agents can access: - orgx://initiative/{id} - Initiative details - ui://widget/decisions.html - Decision card widget - orgx://artifact/{id} - Artifact contents
Pre-built conversation starters:
  • create-roadmap - Scaffold initiative with milestones
  • weekly-recap - Summarize agent activity

​
Next Steps

See Agents in Action

Detailed guide to each agent type

MCP Tools Reference

Complete API for MCP integrations