Capstok — learn by doing

# Assistant — chat + retrieval, low iteration count
def assistant_handler(query, history, user_id):
    chunks = retrieve(query, k=5)
    return llm_call(system=ASSISTANT_PROMPT + chunks, history=history + [query])

# Copilot — embedded in a workflow, latency-critical
def copilot_handler(context):
    # < 500ms p99 target
    return llm_call_streaming(model="claude-haiku-4-5-20251001", messages=[..., context])

# Agent — multi-step with tools
def agent_handler(query, user_id):
    return run_agent_loop(query, user_id, max_steps=6)   # 5-15 seconds typical

# Autonomous — runs without user in the loop
def autonomous_handler(task_id):
    # may run for minutes to hours
    # needs: checkpointing, recovery on crash, progress reporting
    while not task.complete:
        next_action = plan_next(task)
        result = execute(next_action)
        checkpoint(task, result)
        if needs_human_review(result):
            pause_for_review(task)
            return

# Decision tree
# - sync user-facing chat?              -> assistant or agent
# - embedded in another tool?            -> copilot
# - long-running background work?        -> autonomous
# - high-volume cheap queries?           -> probably assistant
# - few queries, deeply researched?      -> agent or autonomous

Run: python3 main.py

Assistant, Copilot, Agent, or Autonomous?

Why this matters

Demo

Try it yourself

References

Chat about this lesson