LLenny's Podcast
← All frameworks
Innovation

The Agency-Control Autonomy Ladder

Ship AI in graduated versions, trading human control for machine agency only as trust is earned.

Difficulty
Moderate
Time to result
~months to results
Steps
3
Confidence
95%

A design principle for AI products that treats autonomy as a dial you turn up over time, not a switch you flip on day one. You start each capability at high human control / low machine agency, then increase agency and reduce control version by version as the system proves reliable. Use it whenever you are tempted to ship a fully autonomous agent immediately.

Origin

Aishwarya Reganti and Kiriti Badam developed this from 50+ AI deployments and their Maven course. It arose because clients, under competitive pressure, kept trying to build end-to-end autonomous agents that became impossible to debug. They contrast it with traditional deterministic software and frame it around the 'agency-control trade-off.'

Core principles

  • 01Handing an agent more decision-making agency means relinquishing an equal amount of your own control.
  • 02Agency must be earned through demonstrated reliability, not granted up front.
  • 03Never ruin the end-user experience while calibrating; only remove the human's control layer.
  • 04Starting small forces you to define the actual problem instead of chasing solution complexity.
  • 05You can constrain autonomy by number of actions, by topic/risk, or by problem complexity.

How to run it

  1. 1

    Start at high control, low agency

    Deploy the AI as a suggestion engine only. It proposes; a human decides and acts. E.g. a support AI suggests answers to human agents, or a coding assistant offers inline completions.

    Pro tip Log everything the human does at this stage — accepted, edited, or discarded — so you get error analysis for free and build a flywheel.

    Watch out Even simple-sounding tasks like ticket routing hide brutal complexity (messy hierarchical taxonomies, dead nodes from 2019); do not assume the first rung is trivial.

  2. 2

    Confirm reliability, then raise one rung

    Once the system performs well and humans rarely override it, increase agency: show answers directly to customers, or generate larger blocks like tests and refactors for human review.

    Pro tip Advance only when calibration stops surfacing new behavior — 'minimizing surprise' is the signal you are ready.

    Watch out Do not jump straight to the ideal V3; teams that leap to full autonomy get buried debugging emergent failures in 100 different ways.

  3. 3

    Add functionality and reach end-to-end autonomy

    Only after lower rungs are stable, expand scope (issue refunds, raise feature requests, apply changes and open PRs autonomously) and let the AI resolve end-to-end for low-risk cases.

    Pro tip Segment by risk: autonomous for MRIs/blood tests, human-in-the-loop for invasive surgery; autonomous for some topics, gated for others.

    Watch out A new data distribution can appear even at V3 (users invent new query types); autonomy is never 'done' — recalibrate when behavior shifts.

In the wild

Customer support agent (OpenAI-scale)

Rather than dumping all help-center articles into an agent, start by having the AI suggest responses to human support agents, capture which suggestions were good or bad, then graduate to showing answers directly to customers, then to issuing refunds and filing feature requests.

Complexity stays controllable and the team builds a labeled feedback loop instead of an unfixable black box.

Coding assistant progression

V1 suggests inline completions and boilerplate; V2 generates larger blocks like tests or refactors for human review; V3 applies changes and opens PRs autonomously.

Trust and reliability are established at each rung before more agency is granted.

Common mistakes

Jumping to V3 (full autonomy) on day one

Because you can't predict user inputs or LLM outputs, a huge multi-step autonomous workflow becomes a debugging and hot-fixing nightmare — the team in the transcript had to shut a product down over uncountable emergent errors.

Obsessing over solution complexity instead of the problem

Starting big lets you skip defining the actual problem; starting small forces problem-first thinking and reveals which slice AI can genuinely solve.

Is it for you?

Best for

Product leaders and AI engineers building agentic features in enterprises where reliability and trust gate deployment.

Not ideal for

Pure research demos or throwaway prototypes where there is no end-user trust to protect and no need to preserve an experience.

From the transcript

every time you hand over decision-m capabilities or autonomy to agentic systems, you're kind of relinquishing some amount of control on your end

10:00

So if you start all with all of this on day one, it's incredibly hard to control the complexity. So we recommend like you know…

15:00

you keep that as is but then remove the amount of control that the human has and there is no single right way of doing…

16:00

From the episode

Why most AI products fail: Lessons from 50+ AI deployments at OpenAI, Google, and Amazon