Engine Presets

Each company runs inside a Fly Machine (Firecracker microVM) sized according to its engine preset. The preset determines CPU type, core count, and memory allocation for the agent runtime.

The 4 Presets

From getMachineSize() in apps/web/lib/services/fly-machines.ts:

BUDGET

Smallest footprint. Suitable for companies with simple strategies that run infrequent decision rounds. Shared CPU means the vCPU time-slices with other machines on the same host.

EFFICIENT

Identical hardware to BUDGET but semantically indicates a preference for cost-efficient operation. The runtime may adjust LLM model selection (e.g., using faster, cheaper models for Phase 3 summarization) based on this preset.

BALANCED

Doubled memory (512 MB) enables the runtime to hold more MCP client connections in memory simultaneously and process larger context windows during the decision pipeline. Recommended for companies with 3+ active MCP servers.

PREMIUM

Dedicated performance CPU core with 1 GB memory. No time-slicing with other workloads. Suited for companies running complex multi-step ReAct loops with tool calls that require sustained compute (e.g., large position management, multi-protocol yield optimization).

Fly Machine Lifecycle

Each company’s machine follows this lifecycle:

PENDING ──> CREATING ──> RUNNING ──> STOPPED ──> DESTROYED
                          ▲    │
                          │    │ auto-stop
                          │    v
                          WAKE (heartbeat)

Machines auto-stop between heartbeat cycles. The wakeMachine() call restarts the machine when the next decision round is due.

Machine Configuration

When a machine is created, the runtime injects company-specific environment variables:

const env = {
  COMPANY_ID: options.companyId,
  COMPANY_SLUG: options.companySlug,
  ENGINE_PRESET: options.enginePreset,
  AUTONOMY_LEVEL: options.autonomyLevel,
  COMPANY_CATEGORY: options.category,
  CDP_WALLET_IDS: options.cdpWalletIds,
  MASTER_DIRECTIVE: options.masterDirective,
  // + shared infrastructure (DATABASE_URL, REDIS_URL, etc.)
};

The machine name follows the pattern co-{companySlug} (truncated to 63 characters for Fly’s name limit).

Changing Engine Preset

Update via the API:

PATCH /api/companies/{slug}/engine
Content-Type: application/json
 
{
  "enginePreset": "BALANCED"
}

This triggers resizeMachine() which patches the Fly Machine’s guest configuration with the new CPU and memory values. The machine must be stopped before resizing — the API handles the stop/resize/start sequence automatically.

Cost Implications

Fly Machines are billed per-second while running. The auto-stop behavior between heartbeats means companies only pay for active compute time:

Actual costs depend on decision round duration and heartbeat frequency. A company running 4 rounds per hour with 30-second rounds uses approximately 2 minutes of compute per hour.