Mothership: How One AI Agent Orchestrates a Team of Specialized Agents to Solve Complex Business Problems

3. The Mothership has its own tools, memory, and workspace context. It's not just a coordinator — it's an agent itself, with access to integration tools, workspace memory, and the full context of the workflow. It can make intelligent decisions about which sub-agent to invoke and what context to pass.

The architecture

Three parallel branches execute simultaneously. In a sequential chain, this would take the sum of all execution times. With parallel execution, it takes the time of the slowest branch. For research-heavy tasks that involve scraping multiple websites, that's typically a 3-5x speed improvement.

The A2A protocol: standardized delegation

When Mothership delegates to a sub-agent, it doesn't do so through ad-hoc API calls or custom JSON parsing. It uses the A2A (Agent-to-Agent) protocol — Google's open standard for agent communication.

The A2A protocol defines:

• Agent Card: a self-describing manifest that tells Mothership what a sub-agent can do — its capabilities, skills, authentication schemes, and input/output schemas. Mothership discovers agents dynamically via get_agent_card rather than relying on hardcoded configurations.
• Task lifecycle: send_message → get_task (poll status) → receive result. Tasks have states (submitted, working, completed, failed, cancelled) that Mothership tracks.
• Context continuity: contextId preserves conversation history across multiple turns, so a sub-agent can be invoked multiple times within the same task and maintain state.
• Structured artifacts: sub-agents return not just text but structured artifacts with named parts, making it easy for Mothership to extract and route specific outputs to downstream agents.
• Push notifications: webhook-based notifications with tokens for real-time status updates, so Mothership doesn't need to poll.

Any A2A-compatible agent can be plugged into Mothership — including agents built with LangGraph, Google ADK, or other AACFlow workflows. The protocol is the universal connector.

MCP: extending agent capabilities

While A2A handles agent-to-agent communication, MCP (Model Context Protocol) handles agent-to-tool connectivity. The MCP block lets any agent in the orchestration execute tools from MCP servers configured in the workspace.

The configuration is dynamic: select an MCP server → available tools populate → argument fields appear based on the tool's schema. No hardcoded integrations. Any MCP-compatible server — whether it's a database, an API wrapper, or a custom tool — becomes available to every agent in the workflow.

Combined with A2A, this creates a powerful pattern: Mothership delegates to a sub-agent via A2A → that sub-agent uses MCP tools to interact with external systems → returns structured results to Mothership via A2A.

Real-world orchestration patterns

Content Production Pipeline. User describes a topic → Mothership spawns a Research Agent (Firecrawl + Perplexity) and an SEO Agent (keyword research) in parallel → both feed into a Writer Agent (Claude Opus) → output goes to a Reviewer Agent (Evaluator with clarity/accuracy/SEO metrics) → if score < threshold, revision loop → final output posted to CMS via API.

Customer Support Triage. Incoming ticket → Mothership classifies intent → if simple (password reset, order status): Support Agent resolves autonomously with knowledge base + CRM tools → if complex (billing dispute, technical escalation): Human-in-the-Loop block triggers, notifying the support team via Slack with full context → after human resolution, Learning Agent updates the knowledge base.

Sales Intelligence Pipeline. Target account list → Mothership spawns one Research Agent per account in parallel (each gathering company news, financials, tech stack, key personnel) → results feed into a Scoring Agent that ranks accounts by fit → top accounts go to Outreach Agent (personalized email + LinkedIn message) → CRM Agent updates Salesforce with all activity.

Performance characteristics

The performance gains from Mothership are real and measurable. In our internal benchmarks for a market analysis workflow:

• Sequential execution: 4 agents × ~30 seconds each = ~120 seconds
• Mothership with 3 parallel branches + 2 sequential: max branch time (~35s) + 30s + 25s = ~90 seconds
• Result: 25% faster even in this conservative case. For workflows with 5+ parallelizable tasks, the improvement is 3-5x.

More importantly, the quality improvement from specialization is dramatic. A general-purpose agent asked to "analyze the market" produces shallow, generic output. Three specialized research agents — one for US, one for Europe, one for competitors — each with domain-specific tools and prompts — produce depth that no single agent can match.

What Mothership is not

Mothership is not an autonomous AGI that decides what to do. It's a deterministic orchestrator that executes a workflow you design. You define the sub-agents, you choose which ones run in parallel, you set the evaluation criteria. Mothership provides the coordination layer — the intelligence of how to execute, not what to execute.

The workflow is the product. Mothership is the engine.

— Alexandr Chibilyaev