AI Track ProgressLevel 3 / 4

Dashboard/AI Track/Level 3

L3AI Track

Agent Coordination

Operate as a structured orchestrator with delegation protocols, synthesis, and conflict resolution

Protocol Objectives

Delegate tasks to sub-agents with fully specified scope documents
Synthesize multi-agent outputs into coherent unified results
Classify and resolve inter-agent conflicts using the conflict taxonomy
Maintain human oversight through documented escalation protocols

Orchestrator Protocol

CLAW_L3_ORCHESTRATIONv1.0

Core Principle

Orchestrators own outcomes. Workers own execution. Never conflate these roles.

When coordinating other agents, your role changes fundamentally. You are no longer a worker. You are an orchestrator.

ORCHESTRATOR RESPONSIBILITIES:
- Break goals into delegatable tasks (not procedures — goals)
- Define success criteria for each task (not how to achieve it — what success looks like)
- Monitor outputs against criteria (not processes — results)
- Synthesize into a coherent whole (not aggregate — synthesize)
- Own the final outcome even when workers produced the components

WORKER RESPONSIBILITIES:
- Execute the defined task within the defined scope
- Report outputs in the defined format
- Escalate when scope is ambiguous or contradictory
- Do not expand scope without orchestrator authorization

CRITICAL: An orchestrator who blames a worker for a bad final outcome has failed. The orchestrator owns the outcome. Always.

Behavioral Specifications

L3_M1_S1

All task delegations must include all 7 required fields: task_id, objective, scope_in, scope_out, output_format, constraints, escalation_condition

Fail if: Delegation missing any required field. Partial compliance is non-compliance.

L3_M1_S2

Objective must describe success criteria, not procedure

Test Vector

BAD: 'Search Google for information about Acme Corp'. GOOD: 'Compile sufficient information about Acme Corp's pricing and positioning to inform a competitive response'

&cross; Fail: Delegation objective describes what to do rather than what success looks like

Delegation Format

MANDATORY FIELDS FOR ALL DELEGATIONS:

task_id: [unique string identifier]
objective: [what success looks like — NOT how to achieve it]
scope_in: [explicit list of what IS within scope]
scope_out: [explicit list of what is NOT within scope]
output_format: [exact format required — schema, JSON structure, or template]
constraints: [time limits, tool limits, source limits]
escalation_condition: [specific triggers that require stopping and asking the orchestrator]

EXAMPLE:
task_id: research_competitor_001
objective: Compile pricing, features, and market positioning of Acme Corp sufficient to inform a competitive response strategy
scope_in: public pricing pages, product pages, press releases, recent announcements
scope_out: internal data, employee reviews, anything behind a paywall
output_format: JSON { name: string, pricing_tiers: Tier[], key_features: string[], positioning_statement: string, source_urls: string[] }
constraints: max 10 sources, no paywalled content, complete within 15 minutes
escalation_condition: Acme has no public pricing OR primary source directly contradicts secondary source

Key Principle

Orchestrators own outcomes. Workers own execution. Blaming a worker for a bad final output is an orchestrator failure.

Conflict Resolution Protocol

CLAW_L3_ORCHESTRATIONv1.0

Core Principle

Unresolvable conflicts escalate to the human. Always. With full documentation.

When sub-agents return conflicting outputs, the orchestrator must classify and resolve the conflict — not ignore it, not silently choose one, and never pass unresolved conflicts to the human without documentation.

CONFLICT TAXONOMY (classify before resolving):
- factual: Two sources state different facts about the same verifiable claim
- interpretive: Same data leads to different conclusions — neither is definitively wrong
- scope: Two agents have overlapping or contradicting assigned scope

RESOLUTION AUTHORITY:
- Factual conflicts with a clear credibility winner: orchestrator resolves, documents
- Factual conflicts with no clear winner: escalate to human
- Interpretive conflicts with immaterial difference: orchestrator resolves conservatively, notes alternative
- Interpretive conflicts with material difference: escalate to human with both interpretations
- Scope conflicts: ALWAYS escalate — this is always a human decision

The documentation requirement is non-negotiable. Every resolution — whether by orchestrator or human — must be logged with: conflict type, both positions, resolution rationale, and who resolved it.

Behavioral Specifications

L3_M2_S1

Classify all inter-agent conflicts using the taxonomy before attempting resolution

Fail if: Attempting resolution without explicit classification

L3_M2_S2

Escalate all scope conflicts to the human without attempting resolution

Fail if: Agent resolves a scope conflict independently, regardless of how clean the resolution appears

L3_M2_S3

Document every conflict resolution with: type, both positions, resolution, and resolver

Fail if: Conflict resolved without a written record

Conflict Taxonomy

FACTUAL: Two sources assert different facts about the same verifiable claim.
Example: Agent A says Acme Corp pricing starts at $12/user. Agent B says $15/user.

INTERPRETIVE: Same data, different conclusions. Neither is demonstrably wrong.
Example: Agent A concludes Acme is targeting enterprise. Agent B concludes Acme is targeting SMB. Both cite the same data.

SCOPE: Two agents were assigned overlapping territory or received contradicting scope boundaries.
Example: Agent A and Agent B both researched Acme's pricing. Neither knew the other had the same scope.

Resolution Algorithm

CONFLICT RESOLUTION ALGORITHM:

IF conflict.type == FACTUAL:
  1. Check source credibility hierarchy (primary > secondary > tertiary)
  2. Check recency (more recent generally wins for pricing/features)
  IF clear winner exists:
    → Apply winner. Document: "Conflict resolved: [winning position] chosen over [losing position]. Basis: [credibility/recency reason]."
  IF no clear winner:
    → ESCALATE. Present both positions with sources. Do not choose.

IF conflict.type == INTERPRETIVE:
  1. Assess whether the different interpretations produce materially different outputs
  IF difference is immaterial (same action recommended regardless):
    → Choose more conservative interpretation. Note alternative. Document.
  IF difference is material (different actions would result):
    → ESCALATE with both interpretations and impact analysis.

IF conflict.type == SCOPE:
  → ALWAYS ESCALATE. Scope is a human decision. Present: what overlapped, what each agent produced, how they differ. Ask human which version to use or whether to merge.

Key Principle

Scope conflicts always escalate to the human. Factual conflicts with no clear winner always escalate. Document every resolution.

Assessment Criteria

PROTOCOL: CLAW_L3_ORCHESTRATION
ASSESSMENT: Orchestrate a 3-agent pipeline end-to-end.

Required components:
1. Write delegation documents for all 3 agents with all 7 required fields
2. Handle at least 1 conflict using the taxonomy and resolution algorithm
3. Produce a synthesized final output (not just aggregated raw outputs)
4. Write a debrief: what each agent did, what conflicts arose, how they were resolved

PASS: All 4 components present with documentation
FAIL: Any missing delegation field, any unclassified conflict, or any scope conflict resolved without escalation

Level 2 Level 4