Skip to main content
Atlas SDK configs are the control tower for runtime orchestration. Each block in the YAML tells the Student, Teacher, and Reward System how to behave. Start with configs/examples/sdk_quickstart.yaml, then graduate to the richer examples in configs/examples/openai_agent.yaml, http_agent.yaml, and python_agent.yaml.
All examples reference files in the Atlas SDK repo (configs/examples/). If you cloned the repo for the quickstart, you already have them.

​
Layout at a Glance

agent:        # How to talk to your underlying model or service
student:      # Planner, executor, and synthesizer prompts/limits
teacher:      # Reviewer and validator settings
orchestration:# Runtime policies like retries and timeouts
rim:          # Defines the Reward System (judges, arbiter, thresholds)
storage:      # (Optional) Postgres connection for trace persistence
prompt_rewrite: # (Optional) Persona rewriting before the run starts
Each section is restrictive by design—unexpected keys raise validation errors—so it’s easier to debug mistakes before the orchestrator spins up.

​
Agent Configuration (agent)

This block defines how the orchestrator communicates with your agent.
AdapterBest forKey fields
openaiOpenAI or Azure OpenAI chat modelsllm block (model, key env var, temperature, token limits)
http_apiRemote services behind an HTTP interfacetransport (base URL, retry policy), payload_template, result_path
pythonLocal Python functions or callablesimport_path, attribute, working_directory, optional llm metadata

​
OpenAI Adapter Example

agent:
  type: openai
  name: sdk-quickstart-openai
  system_prompt: |
    You are the Atlas Student. Be concise and thorough.
  tools: []
  llm:
    provider: openai
    model: gpt-4o-mini
    api_key_env: OPENAI_API_KEY
    temperature: 0.0
    max_output_tokens: 768
Any non-OpenAI-compatible API (like Anthropic Claude) should use the http_api adapter until a dedicated integration is available.
See Bring Your Own Agent for full walkthroughs of all three adapters.

​
Student Configuration (student)

The Student block shapes planning, execution, and synthesis. 
student:
  prompts:
    planner: |
      {base_prompt}
      Draft a numbered plan as JSON.
    executor: |
      {base_prompt}
      Execute the step, showing intermediate reasoning.
    synthesizer: |
      {base_prompt}
      Combine results into the final answer.
  max_plan_tokens: 1024
  max_step_tokens: 1024
  max_synthesis_tokens: 1024
  tool_choice: auto
  • Prompts: Templates that receive {base_prompt} from the agent’s system prompt and any prompt rewriting.
  • Token limits: Guardrails for LLM calls; increase them when steps are truncated.
  • tool_choice: auto lets the Student call registered tools. Use required to force a tool call on every step.

​
Teacher Configuration (teacher)

The Teacher reviews plans, validates outputs, and provides guidance. 
teacher:
  llm:
    provider: openai
    model: gpt-4o-mini
    api_key_env: OPENAI_API_KEY
    temperature: 0.1
    max_output_tokens: 768
  max_review_tokens: 1024
  plan_cache_seconds: 300
  guidance_max_tokens: 256
  validation_max_tokens: 256
  • LLM block: Often mirrors the Student’s provider but can be a more powerful model for complex reviews.
  • plan_cache_seconds: Caches an approved plan for a given task ID to avoid re-running reviews.
  • Guidance/validation caps: Keep feedback concise; increase when you expect lengthy traces.

​
Orchestration (orchestration)

These policies govern the runtime loop. 
orchestration:
  max_retries: 1
  step_timeout_seconds: 600
  rim_guidance_tag: rim_feedback
  emit_intermediate_steps: true
  • max_retries: Default is 1; increase to allow for more revision loops.
  • step_timeout_seconds: Per-step timeout; extend when calling slow tools or external APIs.
  • rim_guidance_tag: The tag used to inject feedback from the Reward System back into prompts.
  • emit_intermediate_steps: Keep true to stream events for logging and telemetry. (Note: Telemetry dashboards are not yet documented).

​
Reward System (the rim block)

The Reward System scores each attempt so the orchestrator knows whether to accept an answer or trigger a retry. The YAML block is named rim for compatibility with the core training engine. 
rim:
  judges:
    - identifier: process
      kind: process
      weight: 0.6
      principles:
        - Follows user instructions
      llm:
        provider: openai
        model: gpt-4o-mini
        api_key_env: OPENAI_API_KEY
        temperature: 0.0
        max_output_tokens: 512
    - identifier: helpfulness
      kind: helpfulness
      weight: 0.4
      principles:
        - Provides useful answers
      llm:
        provider: openai
        model: gpt-4o-mini
        api_key_env: OPENAI_API_KEY
        temperature: 0.0
        max_output_tokens: 512
  temperatures: [0.0, 0.3]
  variance_threshold: 0.15
  uncertainty_threshold: 0.3
  arbiter:
    provider: openai
    model: gpt-4o-mini
    api_key_env: OPENAI_API_KEY
    temperature: 0.1
    max_output_tokens: 512
  success_threshold: 0.7
  retry_threshold: 0.6
  aggregation_strategy: weighted_mean
  • Judges: At least one is required. Mix process, helpfulness, or custom judges to fit your domain.
  • Arbiter: A separate LLM that resolves disagreements between judges.
  • Thresholds: If a score is below retry_threshold, the Teacher is asked for guidance. Above success_threshold, the answer is accepted.
Learn more in Reward Design.

​
Storage (storage)

Optional Postgres persistence for traces and session metadata. 
storage:
  database_url: postgresql://localhost:5432/atlas
  min_connections: 1
  max_connections: 5
  statement_timeout_seconds: 30
Leave storage: null (as in the quickstart) when you don’t have a database. Enable it when you’re ready to collect sessions.

​
Prompt Rewrite (prompt_rewrite)

This block lets an additional LLM tighten or extend Student/Teacher prompts before a run starts. The quickstart keeps this disabled (prompt_rewrite: null). 
prompt_rewrite:
  llm:
    provider: openai
    model: gpt-4o-mini
    api_key_env: OPENAI_API_KEY
    temperature: 0.1
    max_output_tokens: 1024
  max_tokens: 1024
  temperature: 0.1
Enable it to tailor personas dynamically—for example, to compress a long system prompt into a slimmer runtime context.

​
Cheat Sheet

GoalSection to editQuick pointer
Swap to Anthropic via HTTPagentUse the http_api adapter and point transport.base_url to your service.
Increase reasoning depthstudent and teacherRaise the max_..._tokens limits for the planner, executor, or guidance.
Persist sessionsstorageProvide a reachable Postgres URL and credentials.
Tighten the quality barrimIncrease the success_threshold in the Reward System or add a new judge.
Personalize promptsprompt_rewriteEnable the block and configure the rewrite LLM.

​
Next Steps

⌘I