Agentlas
Agentlas
← Back to app
Agentlas OS

Model allocation — spending tokens where judgment matters

Not every step in a task deserves your most expensive model. Agentlas separates who judges that from who enforces it, so a cheap step actually runs cheap and an important one still gets real judgment.

Two roles, one policy each

orchestrator
Plan, route, clarify, verify, synthesize. Anything that's judgment about the work itself.
worker
Execute. The actual doing, once the plan is set.

An unrecognized stage stays on orchestrator rather than silently downgrading — the default direction is toward more judgment, not less.

Who decides the tier

  1. A parent or leader AI looks at each unit of work — its complexity, risk, context size, output size, tools needed, and whether it fans out into multiple sub-tasks — and proposes a cost tier: economy, balanced, or frontier, plus an effort level.
  2. The host checks that proposal against what's actually available: live model inventory, any pins you or your organization set, context/tool support, and cost ceilings.
  3. The final choice is recorded as a receipt — never invented after the fact, and never backfilled with fake zero-token usage before the work has actually run.
INFO · High risk doesn't auto-upgrade the model

A risky step doesn't automatically get forced onto your strongest (most expensive) model. Instead it's flagged as needing independent verification — a separate check, narrower permissions, or your explicit approval — because a stronger model making the same mistake isn't actually safer.

Setting a policy

No vendor lock-in: the engine itself doesn't know model names or hardcode any provider's pricing — your host supplies its live inventory, and you can pin exact models per role.

{
  "orchestrator": { "pinnedProvider": "codex", "pinnedModelId": "gpt-5.6-sol", "maxTier": "frontier", "maxEffort": "xhigh" },
  "worker": { "pinnedProvider": "local", "pinnedModelId": "luna", "maxTier": "economy", "maxEffort": "low" }
}

Set as AGENTLAS_MODEL_ALLOCATION_POLICY_JSON on the MCP server (or its Codex/Claude equivalent). Omitting the worker block just means workers inherit the orchestrator's policy — quality-first by default, never the reverse.

Model allocation — spending tokens where judgment matters · Agentlas docs