Adding Resource Quotas to an Agent (Scala)

Golem provides a distributed resource quota system via golem.host.QuotaApi. Quotas let you define limited resources (API call rates, storage capacity, connection concurrency) and enforce consumption limits across all agents in a deployment.

1. Define Resources in the Application Manifest

Add resource definitions under resourceDefaults in golem.yaml, scoped per environment:

resourceDefaults:
  prod:
    api-calls:
      limit:
        type: Rate
        value: 100
        period: minute
        max: 1000
      enforcementAction: reject
      unit: request
      units: requests
    storage:
      limit:
        type: Capacity
        value: 1073741824  # 1 GB
      enforcementAction: reject
      unit: byte
      units: bytes
    connections:
      limit:
        type: Concurrency
        value: 50
      enforcementAction: throttle
      unit: connection
      units: connections

Limit Types

• Rate — refills value tokens every period (second/minute/hour/day), capped at max. Use for rate-limiting API calls.
• Capacity — fixed pool of value tokens. Once consumed, never refilled. Use for storage budgets.
• Concurrency — pool of value tokens returned when released. Use for limiting parallel connections.

Enforcement Actions

• reject — returns Left(FailedReservation) with an optional estimated wait time. The agent must handle the error.
• throttle — Golem suspends the agent until capacity is available. Fully automatic, no code needed.
• terminate — kills the agent with a failure message.

2. Acquire a QuotaToken

Acquire a QuotaToken once per resource, typically in the agent constructor:

import golem.host.QuotaApi._
 
val token = QuotaToken("api-calls", BigInt(1))

The second parameter is the expected amount per reservation, used for fair scheduling. For simple 1-call = 1-token rate limiting, use BigInt(1).

3. Simple Rate Limiting with withReservation

Use withReservation to reserve tokens, run code, and commit actual usage:

val result = withReservation(token, BigInt(1)) { reservation =>
  callSimpleApi().map { response =>
    (BigInt(1), response)
  }
}

The callback returns Future[(BigInt, T)] where the first element is actual usage. If actual < reserved, unused capacity returns to the pool.

4. Variable-Cost Reservations (e.g., LLM Tokens)

Reserve the maximum expected cost, then commit actual usage:

val result = withReservation(token, BigInt(4000)) { reservation =>
  callLlm(prompt, maxTokens = 4000).map { response =>
    (BigInt(response.tokensUsed), response)
  }
}

5. Manual Reserve / Commit

For finer control, use reserve and commit directly:

token.reserve(BigInt(100)) match {
  case Right(reservation) =>
    val result = doWork()
    reservation.commit(BigInt(result.actualUsage))
  case Left(failed) =>
    println(s"Quota unavailable: $failed")
}

6. Splitting Tokens for Agent-to-Agent RPC

Split a portion of your quota to pass to a child agent:

val childToken: QuotaToken = token.split(BigInt(200))
for {
  childAgent <- SummarizerAgent.newPhantom()
  summary <- childAgent.summarize(text, childToken)
} yield summary

The child agent receives the QuotaToken as a method parameter and uses it for its own reservations. Merge returned tokens back:

token.merge(returnedToken)

7. Dynamic Resource Updates via CLI

Modify resource limits at runtime — changes affect running agents immediately:

golem resource update api-calls --limit '{"type":"rate","value":200,"period":"minute","max":2000}' --environment prod

Key Constraints

• Acquire QuotaToken once and reuse — do not create a new one per call
• All quota amounts are BigInt values
• split traps if childExpectedUse exceeds the parent’s current expected-use
• merge traps if the tokens refer to different resources
• withReservation returns Future[Either[FailedReservation, T]] — Left only for reject enforcement; throttle suspends transparently
• Resource names in code must match the names in golem.yaml resourceDefaults