AoT Harness

CHIP + Atom of Thoughts (AoT) — multi-provider (Anthropic, OpenAI, Google, Mistral, OpenRouter)

AoT Harness

AI-generated

Summary

Perform advanced AI task processing using multi-provider large language models with configurable modes: full AoT decomposition plus QA, AoT only decomposition, Process Analyst for operational case analysis, or webhook relay.

Inputs

  • provider (required) — Select the large language model provider to use for atom execution and question answering.
  • model (required) — Choose the specific model from the selected provider, defaulting to the provider's recommended option.
  • goal — Describe the task or objective for AI processing as a natural language string. Required unless in Process Analyst mode.
  • mode (required) — Choose the operation mode: 'chip' for full AoT with QA, 'aot' for AoT decomposition only, 'process_analyst' for analyzing operational cases, or 'webhook' to call an external Python server.
  • analysisInput (required) — In Process Analyst mode, provide the operational case text (e.g., email, ticket, document excerpt) to analyze. Ignored in other modes.
  • processProfile — In Process Analyst mode, select the profile biasing the analysis focus (generic, email/support, document intake, automotive service).
  • outputLanguage — In Process Analyst mode, set the language for the output fields and draft response (German or English).
  • humanReviewThreshold — In Process Analyst mode, confidence score threshold below which human review is required.
  • includeDraftResponse — In Process Analyst mode, whether to generate a polite draft reply to the operational input.
  • strictJson — In Process Analyst mode, whether to fallback safely upon invalid JSON from the model instead of throwing an error.
  • webhookUrl — In webhook mode, specify the URL of the running aot-harness Python server to forward the request to.
  • providerMix — Optionally enable cost-saving by using a different provider/model for AoT decomposition separate from atom execution.
  • decomposerModel — Select model for the decomposer LLM when provider mix mode is enabled.
  • advanced — Advanced options including QA score threshold, max atoms, max tokens per atom, QA retry behavior, and output language.

Output shape

A single JSON object containing the processed result. Depending on the mode, output fields differ substantially: Process Analyst outputs structured operational analysis with fields like process type, detected entities, urgency, confidence, next best action, and optional draft response. CHIP/AoT modes output the full task result, QA score, success flag, details of atoms processed, usage costs, and provider/model info.

Output always includes provider and model information used, token/cost aggregates, and any warnings or notes like parser errors. Process Analyst mode performs one LLM call and returns analysis and cost. CHIP/AoT modes perform multi-step AoT decomposition, atom solving, and optional QA with retry, aggregating results. Webhook mode sends an HTTP POST to a specified Python server endpoint and relays its JSON response.

Examples

Example 1: Analyze an email support request to identify next steps with structured data and a draft reply.

Set mode to 'process_analyst', provide the operational email text to 'analysisInput', optionally select 'email_support' profile, choose output language, and set human review threshold and draft response preferences.

Example 2: Decompose a complex goal into sub-tasks (atoms), solve them using selected LLM provider and model with quality assurance, returning a final consolidated output.

Set mode to 'chip' (default), provide a natural language 'goal', choose provider and model for atom solving, configure advanced QA parameters and optionally enable provider mix for cost savings.

Links

Discussion