AgentForge Fleet

A governed multi-agent digital workforce that plans, delegates, remembers, executes, and verifies real business operations with Gemini, Google ADK, and Google Cloud.

Comment

Inspiration

Small and mid-sized businesses need AI systems that do more than answer questions. They need digital workers that can coordinate tasks, preserve context, execute authorized actions, and independently verify their work while remaining governed, observable, and accountable.

AgentForge Fleet explores a practical question:

Can specialized AI agents operate as a governed digital workforce instead of a collection of chatbots?

AgentForge Fleet was built by Carriersfy, LLC specifically for the Google All Things Agentic Hackathon.

What it does

AgentForge Fleet is a governed multi-agent system designed to execute and verify real multi-step business workflows.

The fleet consists of three specialized agents:

  1. Mission Coordinator — understands the objective, creates the execution plan, defines acceptance criteria, delegates work, and tracks mission state.
  2. Execution Specialist — executes the authorized plan, uses tools and data sources, produces structured outputs, and captures evidence.
  3. Independent Reviewer — independently validates completeness, checks contradictions, verifies evidence and claims, scores quality and risk, and approves or requests changes.

A deterministic Evidence Gate provides an additional control layer before a mission can be accepted.

The system also supports A2A interoperability, persistent cross-session memory, human approval controls, mission isolation, controlled memory deletion, and auditable execution.

The goal is useful autonomy with accountability.

How we built it

AgentForge Fleet uses Google's agentic and cloud stack:

  • Gemini 3.5 Flash for agent intelligence
  • Google Agent Development Kit (ADK) for multi-agent orchestration
  • Vertex AI for Gemini model serving
  • Google Cloud Run for hosted serverless deployment
  • Firestore for persistent Memory Capability v1
  • Application Default Credentials (ADC) for Google Cloud authentication
  • A2A / Agent Card for agent interoperability
  • FastAPI for the application gateway
  • Deterministic evidence gates for policy, quality, risk, and evidence enforcement

The architecture deliberately separates planning, execution, and independent verification:

Business Request → Mission Coordinator → Execution Specialist → Independent Reviewer → Evidence Gate → Human Approval when required → Final Result

The application is deployed on Google Cloud Run, and the repository contains reproducible testing and deployment instructions for hackathon judges.

Memory Capability v1

A major part of AgentForge Fleet is durable operational memory.

Memory Capability v1 uses live Google Cloud Firestore and demonstrates:

  • cross-session recovery
  • tenant/user/mission scope isolation
  • persistent mission state
  • fail-closed pending approval
  • human-approval resume
  • controlled operational-memory deletion

This enables agents to recover governed mission context across separate sessions without relying solely on the model's conversation window.

Challenges we ran into

The primary challenge was balancing autonomy with control.

Enterprise agents need to perform useful work without requiring human intervention at every step, while preventing unauthorized, insufficiently verified, or hallucinated actions from reaching final execution.

A second challenge was durable context. Long-running agentic work requires persistent operational state without repeatedly placing complete conversation histories into the model context.

A third challenge was maintaining strict separation of responsibilities between agents so that planning, execution, and verification remain independently accountable.

Accomplishments that we're proud of

AgentForge Fleet was designed around accountable digital work rather than conversational AI alone.

The completed architecture incorporates:

  • three specialized and independently accountable agent roles
  • controlled A2A delegation
  • deterministic evidence gating
  • persistent Firestore memory
  • cross-session recovery
  • tenant/user/mission isolation
  • human-in-the-loop approval controls
  • observable and auditable execution
  • Google Cloud deployment
  • reproducible testing

The Memory Capability v1 certification suite completed with:

12 tests passed / 19 warnings / 102.66 seconds

Live Firestore certification also demonstrated:

  • SESSION_A_WRITE: PASS
  • SESSION_B_RECOVERY: PASS
  • CROSS_SCOPE_ISOLATION: PASS
  • PENDING_GATE_FAIL_CLOSED: PASS
  • HUMAN_APPROVAL_RESUME: PASS
  • OPERATIONAL_MEMORY_DELETE: PASS
  • MEMORY_CAPABILITY_V1: PASS

The hosted A2A mission certification completed successfully with all three specialized agents observed and the mission reaching a completed state without failure.

What we learned

Reliable agentic systems require more than powerful models.

They require architecture that makes decisions, actions, evidence, failures, approvals, memory, and recovery observable and governable.

We learned that persistent memory becomes significantly more useful when it is scoped and controlled rather than treated as unlimited conversational history.

We also learned that independent verification and deterministic evidence gates can complement model reasoning by creating explicit boundaries around autonomous execution.

AgentForge Fleet demonstrates how Gemini, Google ADK, Firestore, Vertex AI, and Google Cloud Run can work together as a governed multi-agent digital workforce designed not only to reason, but to perform accountable work.

What's next

The next evolution of AgentForge Fleet will expand its governed agent catalog, tool integrations, observability, evaluation coverage, and long-horizon asynchronous mission capabilities while preserving strict identity, authorization, memory-isolation, and evidence requirements.

Built With

Share this project:

Updates

Submission history