Configuration

RLMConfig

All configuration is managed through the RLMConfig dataclass. Every field has a sensible default.

from fast_rlm import RLMConfig

config = RLMConfig.default()

Fields

Field	Type	Default	Description
`primary_agent`	`str`	(required, no default)	Model used for the root agent. Must be set or `run()` raises.
`sub_agent`	`str`	`primary_agent`	Model used for child subagents. When unset, falls back to `primary_agent`.
`max_depth`	`int`	`3`	Max recursive subagent depth
`max_calls_per_subagent`	`int`	`20`	Max LLM calls a single subagent can make
`truncate_len`	`int`	`2000`	Output characters shown to the LLM per step
`max_money_spent`	`float`	`1.0`	Budget cap in USD — raises an error once cumulative spend exceeds it (see note below; it's a soft cap, not a hard ceiling)
`max_completion_tokens`	`int`	`50000`	Max total completion tokens across all subagents — raises an error if exceeded
`max_prompt_tokens`	`int`	`200000`	Max total prompt tokens across all subagents — raises an error if exceeded
`max_global_calls`	`int`	`∞` (`50` for ACP)	Max total LLM calls across the whole run (root + all subagents, every backend) — raises an error once reached. Counts calls, not tokens, so it's the only budget that works for ACP agents.

Modifying config

RLMConfig is a dataclass — just set attributes directly:

config = RLMConfig.default()
config.primary_agent = "openai/gpt-5.2"
config.max_depth = 5
config.max_money_spent = 3.0
config.max_completion_tokens = 100000
config.max_prompt_tokens = 500000

Using a dict

You can also pass a plain dict if you prefer:

from fast_rlm import run

result = run(
    "Summarize this document",
    config={"primary_agent": "openai/gpt-5.2-codex", "max_depth": 5},
)

Dict values are merged on top of the defaults from rlm_config.yaml.

max_money_spent is a soft cap, not a hard ceiling

The budget is tracked as a single running total across every LLM call in the trace — the root agent, all nested subagents at every depth, and the delegation-confirmation calls. That accounting is complete. But the limit is checked after each call returns, so it can only stop the next call, never calls already in flight.

This matters because batch_llm_query(...) fans out subagents in parallel. All of a batch's calls launch before any of them pushes the cumulative total over the limit, so their tokens are spent (and billed) before the cap trips. Realized spend can therefore overshoot the limit by roughly batch_width × cost-per-call — e.g. a 5-wide batch against a $0.05 cap was observed to reach ~$0.19 before halting. Treat max_money_spent as "stop once I notice I'm over," not a guaranteed ceiling, and set it with headroom below your true limit. Deep/wide configs (max_depth, parallel batches) overshoot more.

The cap is also per run (per process): the running total resets each time the engine starts, so separate invocations — e.g. benchmarking two models — each get their own independent budget. Nothing tracks spend across runs.

Cost tracking depends on your provider

Not all API providers return cost information in their responses. OpenRouter includes cost data, but OpenAI and most other providers do not. If your provider doesn't return cost, max_money_spent won't be able to enforce a budget at all (the check is skipped when cost is null) and cost will show as "Unknown" in the UI. In that case, use max_completion_tokens and max_prompt_tokens to control spending instead — these work with any provider since token counts are always returned.

How config merging works

Defaults are loaded from the bundled rlm_config.yaml
Your overrides (from RLMConfig or dict) are applied on top
The merged config is written to a temp file and passed to the engine

This means you only need to specify what you want to change — everything else keeps its default value.

Model names

fast-rlm uses OpenRouter by default, so model names follow the OpenRouter format: provider/model-name.

Since agents write and execute Python code in a REPL, choose models that are strong at coding. Recommended models:

Model	Name
GPT-5.2 Codex	`openai/gpt-5.2-codex`
Claude Sonnet 4.6	`anthropic/claude-sonnet-4-6`
Gemini 3.1 Pro Preview	`google/gemini-3.1-pro-preview`
MiniMax M2.5	`minimax/minimax-m2.5`
GLM-5	`z-ai/glm-5`

Browse the full list at openrouter.ai/models.

Using a different provider

If you're pointing RLM_MODEL_BASE_URL at a different API (e.g. OpenAI directly), use that provider's model names instead (e.g. gpt-5.2-codex instead of openai/gpt-5.2-codex).