> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mira-app.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# RLM Engine

> How the Recursive LLM Engine writes and executes Python code, observes real output, and iterates until answers are verifiably correct.

The **RLM Engine** (Recursive LLM) approaches problems the way a careful analyst does: by writing Python code to investigate, executing it, observing what actually happened, and refining until the result is verified.

## How it works

```
Question received
      │
      ▼
  Write Python code to investigate
      │
      ▼
  Execute in a local Python REPL
      │
      ▼
  Observe the actual output
      │
      ▼
  Verified? ──No──► Refine reasoning, rewrite code, repeat
      │
     Yes
      │
      ▼
  Return a confident, evidence-backed response
```

Every iteration streams to you in real time via the [REPL Console](/features/repl-console). Every piece of code that ran, every output it produced — visible, auditable, honest.

## Why code execution matters

When a language model answers a quantitative question without executing code, it generates a statistically likely answer — which may or may not be numerically correct. RLM removes that ambiguity: the Python interpreter computes the answer, not the model's internal statistics.

**What this means in practice:**

* Calculations are computed, not inferred
* Cross-referenced data is matched by code, not by pattern matching
* Statistical results come from actual `code` calls, not approximations
* If the code raises an exception, the engine sees the actual error and corrects its approach

## Iteration limit

RLM iterates until either:

* The answer is verified (the code runs successfully and produces a result the model judges correct)
* The **max iterations** limit is reached (default: **30**)

At the limit, RLM returns its best answer based on all iterations completed, along with a note that it did not reach a verified conclusion.

## Execution environment

Python code executes inside the bundled Python virtualenv (`mira-venv/`) on your local machine. There is no sandbox — the code runs with the same OS user permissions as the MIRA process:

* It can read and write any file your OS user can access
* It can make outbound network calls
* It can import any library installed in the virtualenv

This is intentional: unrestricted execution is what allows RLM to work with real local data and external APIs. The tradeoff is that you should only run queries you trust — or review the generated code in the REPL Console before it executes.

## Provider support

| Provider        | Models (examples)                                                       |
| --------------- | ----------------------------------------------------------------------- |
| **AWS Bedrock** | eu/us.anthropic.claude-sonnet-4, claude-3-5-sonnet-v2, claude-3-5-haiku |
| **Anthropic**   | claude-3-5-sonnet, claude-3-5-haiku, claude-3-opus                      |
| **OpenAI**      | gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-3.5-turbo                         |
| **Ollama**      | llama3.2, mistral, qwen2.5-coder, deepseek-r1, and any local model      |

Default: **AWS Bedrock** with `eu.anthropic.claude-sonnet-4-20250514-v1:0`

## Concurrency

RLM supports up to **3 concurrent sessions** by default (configurable). If you send a query while 3 sessions are already in flight, MIRA shows a concurrency-limit error and queues the request.

## Configuration reference

See [RLM Settings Reference](/configuration/rlm-settings) for all configurable parameters with defaults and allowed ranges.

## Best used for

* Complex data analysis — "analyse this CSV and find the top 3 factors driving churn"
* Multi-step calculations — financial modelling, statistical tests, simulations
* Cross-referencing — "find every row in this dataset that contradicts this policy document"
* Anything where being provably correct matters more than being fast

<Tip>
  If speed matters more than verifiability, switch to the [Native Agent
  Engine](/core-concepts/native-agent-engine) — it reaches conclusions faster for open-ended
  questions where code execution isn't needed.
</Tip>