Agentic AI Patterns

Course Content

Agentic AI Patterns

9 sections · 50 lessons

What are autonomous agents vs. assisted agents, and when is each preferred?

What you need to know

The four questions

  • Blast radius: what is the worst outcome of one wrong action?
  • Reversibility: can it be undone cheaply and quickly?
  • Verifiability: is there an automatic check (tests, reconciliation, schema)?
  • Volume: would human review become a bottleneck?

Lean autonomous

  • Log and alert triage
  • Data enrichment and tagging
  • Drafts that a human sends later
  • Holding a booking that expires in 24 hours

Lean assisted

  • Payments and refunds above a limit
  • Emails to customers or vendors
  • Production changes and deletions
  • Clinical, credit or claim-denial decisions

Autonomy belongs to actions, not agents

The same agent can be autonomous for reads and assisted for writes. Tag every tool with a tier, and let code, not the model, enforce it:

Python
TIERS = {    "get_quote":        "auto",       # read-only    "hold_booking":     "auto",       # reversible, expires in 24 h    "create_po":        "approve",    # commits money    "delete_vendor":    "forbidden",  # never from an agent}def gate(tool, args, approval_limit=50_000):    tier = TIERS.get(tool, "forbidden")            # unknown tools are denied    if tier == "approve" and args.get("amount", 0) <= approval_limit:        tier = "auto"                              # small POs may run alone    return tier

gate("create_po", {"amount": 32000}) returns auto; with 480,000 it returns approve; any unknown tool returns forbidden. The default is deny.

The promotion ladder

  1. Ship every write action as assisted.
  2. Record approve, edit and reject rates per action type.
  3. Promote an action to autonomous when its approval rate is high and edits are trivial, for example 98% approved unchanged over 500 cases.
  4. Demote it automatically if its error rate rises.

The approvals also become labelled training and evaluation data at no extra cost.

A real-life example

A procurement agent raises purchase orders for office supplies across 40 branches.

  • Month 1: every PO is assisted. Buyers approve 2,100 POs; 96% are approved unchanged. Almost all edits are on POs above Rs 50,000, usually a changed vendor.
  • Month 2: POs up to Rs 50,000 from approved vendors become autonomous. That covers 78% of volume. Buyers now review about 15 POs a day instead of 70, and actually read them.
  • A month later, a vendor's catalogue price feed breaks and shows prices 10 times too low. The agent's "price differs from last order by over 40%" check flags it, and autonomous POs for that vendor fall back to assisted automatically.

Follow-up questions to expect

  • "Isn't assisted mode slower?" — Yes, and that is the price of safety on risky actions. Keep it only on the actions that need it, so humans review fewer items more carefully.
  • "Can the model decide when to ask a human?" — It can request escalation on low confidence, which is useful. But mandatory gates must be enforced in code, because a confused model will not know it is confused.
  • "What about reversible but high-volume actions?" — Make them autonomous with an easy undo, a visible audit log and rate limits.