Human-Centred Agentic Intelligence (HAI): A Three-Layer Framework for Designing Human-Agent-System Journeys

Research
1

Linkedin
Linkedin1584×850 50 KB

Hi all :waving_hand:

I’m Anandakumar, an independent researcher from London. I’ve just published
an open-access whitepaper on a design framework for agentic AI systems.

:hugs: Hugging Face: HAI Framework - a Hugging Face Space by anandartin

Full paper (free PDF): Human-Centred Agentic Intelligence (HAI): A Three-Layer Framework for Designing Human-Agent-System Journeys

THE PROBLEM

Teams building agentic AI face two different design challenges — and most
frameworks conflate them:

• Task-level: How does the agent behave at each STEP of a product interaction?
• Lifecycle-level: How does agent behaviour evolve across a customer’s full
RELATIONSHIP (Awareness → Purchase → Retention → Advocacy)?

These need different vocabularies — but share the same underlying infrastructure.

THE HAI FRAMEWORK

A Dual-Mode, Three-Layer design matrix:

Three shared layers:
→ Human Layer — what the person sees, does, and feels
→ Agent Layer — how the agent reasons, acts, and handles failure
→ System Layer — data, infra, protocols, and governance

Two modes:
→ User Journey Matrix (Step-Based) — for product/task design
→ Customer Journey Matrix (Stage-Based) — for lifecycle/CX design

Agent Mode Taxonomy per phase:
→ Assistive (human controls) / Advisory (agent suggests) / Autonomous (agent acts)
Autonomy should increase with trust — not be set globally.

ALSO INCLUDES

• 7-type Goal Failure Response Protocol with mode-specific recovery paths
• Logical Handover Framework (5-state model + 9-element context package)
• Multi-Agent Role Taxonomy (Orchestrator / Specialist / Verifier / Liaison)
• MCP + A2A Protocol Layer integration per step/stage
• EU AI Act risk classification at step and stage level
• Excel template with dropdowns — free download in the Zenodo record

QUESTIONS FOR THE COMMUNITY

  1. Does the step-based vs. stage-based distinction match challenges you’ve faced?
  2. Any gaps in the Assistive/Advisory/Autonomous taxonomy from your experience?
  3. How are your teams handling EU AI Act compliance at the step/stage level?

Happy to discuss — especially on failure handling and Agent Mode design.

License: CC BY-NC-ND 4.0 | Anandakumar Muniasamy Pothiraj, London

2

It is my pleasure to welcome your first post with this first reply.
Welcome.

License: CC BY-NC-ND 4.0 | Anandakumar Muniasamy Pothiraj, London

Ah this is the first time I have come accros this " The CC BY-NC-ND 4.0 license by Anandakumar Muniasamy Pothiraj (London) allows sharing, copying, and distributing the material in any medium/format for non-commercial purposes only, provided proper attribution is given. No derivatives (modifications/adaptations) are allowed, and you cannot use the work commercially."

Thanks..

Are you aware of the datatype “Dynamic Unary?” Yeah, that’s mine..

Again welcome.

-Ernst

1 Like
3

Thanks Ernst, appreciate the welcome — good to be here.

The CC BY-NC-ND 4.0 licence is deliberate. HAI is a research framework with specific conceptual contributions — the three-layer matrix, the dual-mode architecture, the agent mode taxonomy — and at this stage I want attribution preserved and adaptations coordinated with me rather than forked freely. Non-commercial is the right default while the work is still establishing itself.

“Dynamic Unary” isn’t a term I’ve encountered — do you have a link or post where you’ve written it up? Happy to take a look.

-– Andy

4

Yeah I was into attempting to compress random binary back in the days.
That dataset is based on the RAND A Million Random Digits with 100,000 Normal Deviates | RAND

I presented my best effort in 2014 : Ernst03 (Ernst Berg) · GitHub
So kind of a crazy thing to dedicate to as a workingman but nonetheless useful today in this conversation.

Mark Nelson, of data compression lore, was the lead voice in debunking shysters claiming to compress random data. This was before the proper Internet; on Usenet. He made the put-up-or-shutup challenge and I found it the “Cat’s Meow” for my interests.
It has been the dataset of my life for all my information science quests. I was a layman; a man of a manual labor career.

However for you, if you have not been advised of limit cycles this is informative.
This “Dynamic Unary Encoding” simply counts runs of same parity in a finite binary and reports the counts in terminated unary encoding.

I am aware that now, that LLM-AI is here, that I can draw upon that wealth and present a better paper in the future.
However, I can answer questions presented.
It simply is a datatype of a dynamical nature.

-Ernst