Skip to main content

​
Intents

An intent is a natural-language description of what you want. a21e parses it into structured fields — primary task, success criteria, constraints, risk tolerance — then routes it through the execution pipeline. 
{
  "input": "Refactor this function to use async/await instead of callbacks",
  "auto_execute": true,
  "model_tier": "standard"
}
Intents can be submitted via the RPC endpoint, the OpenAI-compatible endpoint, or the Huddle web interface.

​
Prompt synthesis

a21e doesn’t just forward your message to a model. It compiles an optimized prompt by:
  1. Technique selection — matching your intent to curated prompt strategies (chain-of-thought, few-shot examples, role prompting, etc.)
  2. Context injection — adding your workspace preferences, memory, persona directives, and repository context
  3. Constraint enforcement — applying org-level policies and user preferences
  4. Quality gating — scoring the output and capturing feedback for future improvement

​
Model tiers

TierBehaviorBest for
autoa21e selects the best model based on task complexityMost requests
economyFastest, lowest costSimple lookups, formatting
standardBalanced quality and speedCode generation, writing
premiumHighest capability modelsComplex analysis, architecture

​
Deliberation

For complex decisions, a21e can run multi-model deliberation:
  1. Multiple models each generate a plan independently
  2. Each model critiques the other models’ plans
  3. A consensus vote determines the best approach
  4. The winning plan is optionally executed
This surfaces disagreements and produces higher-confidence results than any single model.

​
Credits

1 credit = 1 enhancement — one pass through the prompt synthesis and execution pipeline.
  • Managed mode: credits cover prompt engineering + LLM inference
  • BYOK mode: credits cover prompt engineering only (you pay the LLM provider directly)
Credits are consumed from subscription balance first, then add-on balance.

​
Memory

a21e maintains persistent memory across sessions:
  • Corrections — “I prefer snake_case, not camelCase” is remembered for future requests
  • Preferences — language, framework, coding style, verbosity
  • Context — project details, architecture decisions, team conventions
Memory is scoped to user and optionally to organization. It’s automatically retrieved and injected when relevant.

​
Workspaces

A workspace binds together:
  • Repository context — connect a GitHub repo for code-aware responses
  • Preferences — default model tier, risk tolerance, verbosity
  • Persona — custom system instructions for all responses
  • StylePrint — extracted design tokens for consistent output style
Workspaces can be personal or organization-level, with enforced org policies overriding personal preferences.