Guides

Fleet Deployment

Run stacks, stage rollouts, and manage cost, tenancy, and recovery.

Prompt pack
M-10
Source material
  • maple/docs/tutorials/operations.md
  • maple/docs/staged-rollout-checklist.md

Fleet Deployment

Fleet is MAPLE's orchestration layer for running many governed agents as a system rather than as a collection of disconnected demos. In the current repo, the operational control surface is PALM, and the stack-definition surface is the maple-fleet-stack crate.

Current operator workflow

Start PALM:

cargo run -p maple-cli -- daemon start --foreground

Then use the PALM CLI:

cargo run -p palm -- status
cargo run -p palm -- spec list
cargo run -p palm -- deployment list
cargo run -p palm -- playground status

The currently exposed PALM command groups are:

  • spec
  • deployment
  • instance
  • state
  • health
  • events
  • playground

That is the real operator surface today.

Stack topology

maple-fleet-stack already implements a Docker Compose-like stack definition for agents, but it is currently a Rust crate surface rather than a finished maple up CLI.

maple-stack.yml

name: support-stack
version: "1.0"

services:
  support:
    agent_ref: "myorg/agents/support:1.0.0"
    replicas: 3
    environment:
      MAPLE_TENANT: "acme"
    depends_on: []

  evaluator:
    agent_ref: "myorg/agents/support-eval:0.4.0"
    replicas: 1
    depends_on:
      - support

The implemented stack schema supports:

  • agent_ref
  • replicas
  • environment
  • depends_on
  • resources
  • health_check

What is "Docker-compose-like" here

maple-fleet-stack provides:

  • YAML parsing for multi-service agent stacks
  • dependency validation
  • topological sort for startup and teardown ordering
  • stack and service lifecycle state tracking

This is the MAPLE equivalent of a compose file for agents. The final maple up, maple down, and maple ps UX is not wired into maple-cli yet.

Rollout strategy

  • Start with a shadow or canary slice
  • Compare behavior, not only liveness
  • Promote only after Guard and eval signals stay inside threshold
  • Keep rollback artifacts and receipts immediately accessible

Topologies

Local

Single operator, single machine, PALM daemon, and optional Ollama. Best for iteration.

Team

Shared PALM environment with Postgres, playground backends, and basic observability. Best for staging.

Enterprise

Multiple deployments, stronger tenancy controls, managed persistence, and external orchestration. Best for production estates.

Air-gapped

Private registry plus mirrored packages and models. Best for sovereign environments.

Production checklist

  • Define cost budgets per tenant or service
  • Separate approval rules for high-risk actions
  • Retain receipts long enough for audit needs
  • Test rollback with real package versions
  • Prove backup and restore before launch