Fleetwright

Inspiration

Most multi-agent systems stop at orchestration: a workflow graph, a prompt bundle, or several agent classes. Enterprises need more. They need explicit authority, policy enforcement, durable state, human control, recovery, observability, evaluation, and evidence that survives beyond a single model response.

Fleetwright treats the agent organization itself as something that should be compiled.

What it does

Fleetwright is an LLM-driven scenario-to-fleet Creator backed by a deterministic compiler. A user describes an enterprise requirement in natural language, and the Creator proposes roles, collaboration topology, communication protocols, typed tool contracts, knowledge predicates, organizational norms, human authority boundaries, and verification criteria.

That proposal becomes a strict blueprint and then the canonical intermediate representation H = : knowledge, organization, agents, communication, environment and tools, memory, and verification.

Trusted code—not the model—owns canonical identifiers, schemas, paths, source emission, provenance, and safety checks. LinkML, SHACL, typed packs, and contract validation reject invalid organizations before emission. The result is a complete, topology-specific Google ADK harness with model-driving agents, dynamic context, governed tools, sessions, memory, inter-agent communication, stop/resume behavior, human authority, evaluation, deployment contracts, and recovery seams.

Every build produces content-addressed artifacts including fleet.lock, plan.json, roles.json, ir.jsonld, and an output contract, making the generated fleet reproducible and auditable.

How we built it

Fleetwright separates architectural reasoning from trusted compilation:

1. A Creator model interprets the enterprise scenario.
2. A strict blueprint captures its architectural decisions.
3. Trusted Python code materializes the canonical fleet specification.
4. LinkML and SHACL validate structural and organizational semantics.
5. A deterministic emitter generates a complete Google ADK 2.7.1 runtime.
6. The generated fleet applies policy, authority, evaluation, persistence, and evidence contracts at runtime.

The local Creator demo uses Gemma 4 E4B IT. The competition deployment is configured for Gemini 3.5 Flash through Vertex AI and fails closed for non-Gemini providers, older Gemini generations, or silent fallback.

The Control Tower provides a local interface for creating, compiling, comparing, auditing, and running generated fleets. Human approval is represented by a non-LLM pause/resume gate, so a requesting agent cannot approve its own consequential action.

Google Cloud deployment and observed evidence

At submission time, eight authenticated Cloud Run role and broker services were observed Ready in europe-west4. Recent Cloud Logging entries showed healthy endpoints, successful maintenance calls, operational metrics, valid integrity checks, and Google ADK/Vertex/Gemini activity. A real managed job was accepted and dispatched by the deployed coordinator.

We preserve the evidence boundary precisely: the managed job has not yet produced a terminal receipt, so we do not claim completed end-to-end managed qualification. Terminal execution, durable readback, and the full LIVE-VERIFY qualification remain pending. The Control Tower is loopback-only; no public hosted application is claimed.

Challenges

The hardest challenge was preventing “agent fleet” from becoming another name for a workflow. Importing an agent class is not enough: the emitted product must close the real model, session, tool, memory, authority, recovery, and verification loops.

We also had to keep model creativity and system authority separate. The LLM decides architecture, while trusted code owns invariants and fails closed when a proposal violates them.

Finally, we treated infrastructure definitions, local tests, deployed resources, and terminal managed evidence as different proof classes. This made the submission more honest and the system easier to audit.

Accomplishments

• A complete LLM-driving Google ADK harness rather than a static agent export.
• Seven topology-specific, fail-closed organization builders.
• Deterministic compilation with content-addressed provenance.
• Human authority that cannot be self-approved by an autonomous role.
• Typed tools, policies, memory, communication, evaluation, recovery, and observability in one generated system.
• Eight authenticated Cloud Run services observed running, with recent operational logs.
• A real managed job accepted and dispatched while terminal qualification remains honestly marked pending.

What we learned

Governance cannot be added as a prompt after agents are created. It must shape the vocabulary, topology, tools, protocols, authority, persistence, and evidence model from the beginning.

A compiler boundary is valuable for agentic systems: models are strong architectural collaborators, while reproducibility and enforcement belong to deterministic code. Most importantly, deployed infrastructure is not the same as qualified managed execution. Evidence should say exactly what was observed—and stop there.

Other data sources

The demonstrated invoice-compliance scenario uses synthetic, de-identified test data only. No customer or production data is included.

What's next

Next we will complete the terminal managed execution and independent readbacks, exercise negative IAM controls, bind the exact Gemini model invocation to the resulting trace, and finish the full LIVE-VERIFY qualification. A public hosted Control Tower will only be introduced after its authentication and disclosure boundaries are ready.

Reproducibility and disclosure

Development began on August 19, 2026, within the hackathon build period. The repository contains step-by-step local setup, compilation, testing, architecture, and live-verification instructions.

No pre-existing proprietary or previously published author code was reused. Claude Code and Codex were used as permitted AI coding assistants. Third-party open-source components retain their own licenses and are declared in the repository's dependency and lock files.

Built With

• google-adk