Rasa Hero. NLP/NLU, Chatbots, Voice, Conversational UI/UX, CX Designer, Developer, Ubiquitous User Interfaces. 

https://cobusgreyling.medium.com
https://www.eliza.community
https://www.cobusgreyling.me

CX requires four ingredients:

1. Continuity: Vertically in one Medium & Horizontally across Mediums.
2. Intelligent Directed Dialog. 
3. Contextual Awareness. 
4. Intent Discovery

Cobus Greyling holds 4 tertiary qualifications, including a Post-Graduate diploma from the University of Pretoria in GIS. 

Over a period of more than 20 years he has visited 19 countries, and has worked in 5 countries, with 8 GSM networks, and more than 18 companies. Being involved with self-service in 14 languages on various platforms relating to IVR (inbound, outbound, ASR and Speech Synthesis), Voice Biometrics, Conversational Interfaces (Amazon Echo, Google Home, Chatbots), Cognitive Computing, Linguistic Analysis and Ambient Computing. 

Cobus has experience in the areas of training, development, project management, and executive overviews and presentations. 

Cobus also has experience in GIS (Geographic Information Systems), understanding spatial analysis etc.

Countries worked in:
South Africa (Various companies), Lesotho (Vodacom), Tanzania (Vodacom), DRC Congo (Vodacom), Nigeria (VMobile, Celtel), Finland (Tecnotree)

Countries visited:
United States, England, Congo Brazzaville, Namibia, Botswana, Zambia, Zimbabwe, Egypt, France, Netherlands, Switzerland, Belgium, Germany and Italy.

Yesterday, I watched  Andrej Karpathy's YC talk—and wow, it was packed with insights. But one concept hit me hardest: his "autonomy slider" for AI.

It's a simple yet brilliant idea, dial autonomy up or down based on the context and use case. No more one-size-fits-all AI Agents. 

For me this was such confirmation of something I've seen evolving all year—the shift from rigid AI Agents to agentic workflows on a full spectrum, blending varying degrees of AI independence with human oversight.

YouTube. com/watch?v=LCEmiRjPEtQ

Cobus Greyling

NLP/NLU/LLM, Chatbots, Voicebots, CCAI, Ambient Orchestration, Ubiquitous User Interfaces

at the intersection language & AI

Earlier this year I spoke in Cologne about the ability of AI Agents to decompose complex task into a sequence of subtitles. 

But also, added this, the need for an agency slider, where the level of agency is assigned to a task or a process can be sent by the user.

 There has been some exciting research this year on the Human and AI collaboration spectrum and how tasks can be graded according to the spectrum. 

Kore.ai

The future isn’t just AI Agents… it’s entire Agentic Organisations.

I’ve been writing about this shift for a while: from standalone “AI Agents” to Agentic Workflows — lightweight, interconnected meshes of specialised agents that collaborate to solve complex problems.

Platforms like LangChain (one of the earliest to truly embrace this vision), OpenAI’s Swarm framework, Kore.ai, and others are already building the foundations for this new paradigm.

But Microsoft Research just took it several steps further with the concept of Agentic Organisations.

Imagine an entire company (or department) structured as a dynamic mesh of AI agents:

- An “Organiser” agent triages the goal and dynamically forks sub-tasks

- Dozens or hundreds of specialised “Worker” agents run in parallel (research, code, analyse, verify…)

- Results are asynchronously merged into a coherent, high-quality output

Their new AsyncThink paradigm reduces inference latency by ~28% vs traditional parallel approaches while actually improving accuracy on hard reasoning tasks.

They’re already exploring two scaling axes:
1. Quantity — from 10 → 100 → 1,000+ concurrent agents
2. Diversity — domain-specialised agents + unique tool access

Most exciting? The Human-in-the-Loop evolution:

- Human-as-Organiser (you fork tasks to AI workers)
- Human-as-Worker (AI forks judgment calls to you)
- Collaborative pre-planning between human & AI

We’re moving from AI as a tool → AI as a colleague → AI as an entire organisational layer.

The hybrid Human-AI Agentic Organisation might be the most powerful intelligence model we’ve ever built.

https://lnkd.in/d3PcdgJP

AI for Work & the Skill Difficulty Continuum! How AI can upskill workers,
bridge the performance gap between high and low-skilled employees, and
create complementary performance gains. 

Key phases include interventions & extensions, merging benefits, and the compression effect—leading to sharp jumps in job success.

Kore.ai 
Rashmi Adwani

OpenAI Customer Service AI Agent

A While back OpenAI dropped a GitHub repo with a slick demo of an AI-powered Customer Service Agent interface.

https://lnkd.in/dFUpcuEM

The Agent view has a pane with with the agents which are available, the guardrails and the context which is being managed. 

At the bottom is the conversational logs. 

Obviously on the right of the screen is the conversational UI, or the customer view. 

So think of left as the organisational view, and the right as the customer facing portion.

Here the conversation has unfolded as you can see on the right, and the runner output or logs are on the left.

Nathan Schlaffer, MBA and I again collaborated on creating this content, covering "The Decline of AI Agents and Rise of Agentic Workflows".   

Not long ago, fully autonomous AI agents were touted as the ultimate breakthrough, promising to tackle complex, multi-step tasks with zero human input. 

But as enterprises discovered their limitations—like unpredictability, inefficiency, and governance gaps—a more reliable path emerged: agentic workflows. 

Structured systems harness specialized agents, conditional logic, and human-in-the-loop checkpoints for transparency, scalability, and real-world results.

We poured serious effort into the narrative, examples (like RAG systems and Microsoft's AsyncThink) and visuals to make it both insightful and engaging. 

Hope you find it informative!

Dive into the full piece here: https://lnkd.in/dF_ZC6bK

Michael Kropidlowski
Marcel Korst
Abhijit Mhetre
Rashmi Adwani
Kore.ai

Enterprise AI Product Marketing Leader | Fueling sustainable revenue growth with effective B2B GTM strategies and messaging

CMO, Kore.ai

Global Marketing Executive in B2B Technology

EVP Kore.ai | Strategic B2B Marketing | Entrepreneurship | Public Speaking

Driving Social at Kore.ai |Host @TheCoolCouchPod

Kore.ai

Chief Evangelist: AI Agents & Agentic Workflows

Sri Vishnu

OpenAI Customer Service AI Agent, a while back OpenAI dropped a GitHub repo with a slick demo of an AI-powered Customer Service Agent interface.

This GitHub repo showcases an airline customer service AI Agent built on OpenAI’s Agents SDK.

It features five AI Agents collaborating, led by a Triage Agent that delegates tasks to four specialised sub-agents.

Built-in guardrails prevent off-topic chats and jailbreaks.

Handles conversation context and memory for seamless interactions.

GitHub repo: https://lnkd.in/dfubrRR9

Goodbye AI Agents, Hello Agentic Workflows (Research from Microsoft)

Agentic Workflows are made up of smaller, lightweight AI Agents which acts as nodes in networks of agentic workflows…

I have written extensively about it before, the shift from AI Agents to Agentic Workflows.

In short, an Agentic Workflow is a workflow or mesh of nodes where each node is represented by an AI Agent.

But it is a more lightweight AI Agent than what traditionally was considered as an AI Agent.

And this pattern is forming the foundation of Agentic App building platforms like LangChain, OpenAI Agent Builder, Kore.ai AI for Process and more.

In fact, I would argue that LangChain is one of the first companies to have this vision.

A recent study from Microsoft took things a bit further, and I must say, it opened by eyes to what the future might look like…

Microsoft envisions Agentic organisation, where AI Agent nodes connect via a complex mesh (Agentic Workflows) solve complex problems by working collaboratively and concurrently.

Agentic workflows have the capability to split tasks and run processes in parallel, and bring it all together at the end.

In a recent article I shared a working Python notebook which walks you through the process.

FlowMind Is An Automatic Workflow Generator
The FlowMind research was conducted by JPMorganChase to use LLMs to overcome the rigidity of RPA. FlowMind leverages a prompt template to ground LLM reasoning via APIs.

Data integrity and confidentiality is maintained while generating automated workflows. The effectiveness of FlowMind is demonstrated with success in workflow generation, user interaction & feedback.

FlowMind reminds much of Autonomous Agents as we have come to know them. In particular LangChain Agents and LlamaIndex Agentic RAG comes to mind.

FlowMind & Agents...

In both of these implementations the agent creates a sequence or chain of events (analogous to automatic robotic process creation) which is executed.

The only difference in this regards is that FlowMind creates the process or flow as a Python function. Which it then in turn executes to deliver the answer.

Prompt Recipe & Prompt Template

At the heart of agents as we have come to know them, is a prompt template.

FlowMind makes use of something called a Prompt Recipe, which is shown below. The prompt recipe as JPMorganChase calls it, is in essence a template.

In his YC talk, Andrew Ng distills what actually separates winning startups from the rest:

- Relentless execution speed
- Concrete, specific ideas (not vague visions)

His framing of vague vs. concrete ideas is so good:

“Vague ideas are almost impossible to criticise and therefore almost always sound good.

Concrete ideas, on the other hand, stick their neck out—they’re easy to attack, so the ones that survive scrutiny tend to be much stronger.”

The takeaway: if your idea still feels bulletproof after everyone has taken free shots at it, you’re probably onto something real.

(Original talk clip for context: https://lnkd.in/dEhEVaSV — the relevant part starts around 12:40)

🚀 Goodbye Standalone AI Agents — Hello Agentic Workflows!

New research from Microsoft confirms what many of us have been saying: the future isn’t one big monolithic AI agent trying to do everything… it’s Agentic Workflows — networks of smaller, lightweight AI agents working together as nodes in a collaborative mesh.

Key ideas from Microsoft’s vision:
- Each node is a focused, lightweight AI agent
- Agents communicate & collaborate in complex workflows
- Tasks are decomposed, executed in parallel, and reassembled
- This leads to truly scalable, resilient “Agentic Organizations”

We’re already seeing this pattern dominate modern agent-building platforms:
• LangChain (one of the earliest to fully embrace this paradigm)
• OpenAI’s upcoming agent tools
• Kore.ai’s process-first approach
• And many more emerging frameworks

I’ve been writing about this shift from “single hero agent” → “agentic workflow teams” for a while now, and Microsoft’s latest paper just took it to the next level. It’s eye-opening.

Want to see a real example? 
I recently shared a fully working Python notebook that implements a multi-agent workflow end-to-end → check it out here: https://lnkd.in/g4afUEA2

Full post: https://lnkd.in/d3PcdgJP

Creating context for individual, handcrafted cases is straightforward, but the real challenge lies in scaling it up — and developing mechanisms to do so effectively.

Context scaling involves two key aspects that determine how well systems can handle and process contextual information.

The first is length scaling, which tackles the technical hurdles of dealing with extremely long inputs.

This means expanding context windows from thousands to millions of tokens, while keeping the understanding coherent across long stories, documents, or conversations.

It relies on advanced attention methods, memory techniques, and model designs to manage this extended information without losing track.

https://medium. com/@cobusgreyling/context-engineering-a34fd80ccc26

So good to see Kore.ai part of the Microsoft Agent 365 Ecosystem. 


Michael Kropidlowski 
Christie Dahmen  
Abhijit Mhetre

Corporate Vice President at Microsoft | Board Trustee, King County Library System

Introducing Agent 365 – a new agent control plane to help every organization scale AI agents.

Agent 365 is the control plane for AI agents. Whether your agents are created with Microsoft platforms, open-source frameworks, or third-party platforms, Agent 365 helps you deploy, organize, and govern them securely.

And the excitement is real. Our partners are already getting on board, bringing their innovation and momentum to Agent 365’s fast-growing ecosystem. Broader ecosystem of agents such as Adobe, Databricks, Cognition, Genspark, Glean, Kasisto, Manus, NVIDIA, n8n, SAP, ServiceNow, Workday, and more.

Open-source agents published on GitHub for Anthropic, Crew.ai, Cursor, LangChain, OpenAI, Perplexity, and Vercel.

https://lnkd.in/ghAPj6rN

Deep Research has emerged as truly one of the break-out use-cases for GenAI...with a number of frameworks being made available and open-sourced. 

Most notably from Salesforce, NVIDIA and OpenAI. Just three observations...

1.  Deep Research is not vulnerable to latency like customer facing Conversational UIs. And Automates time-consuming tasks.

2. Deep Research is not only limited to personal use and academic research.  It is successfully used for large scale auditing of financial records, anomaly detection and risk assessment in banking, predictive maintenance in manufacturing etc.

3. A deep research pipeline is something you can easily build yourself, in the article attached, I share a Colab notebook, based on OpenAI's research API architecture.  it is a pipeline that orchestrates multiple models to achieve an optimal research framework.

Earlier this year I spoke at the World Summit AI in SF,  in one of the breakout sessions. 

What really surprised me was how many people attended to talk, it was packed. 

And secondly how many people stayed behind to have a talk afterwards. 

It’s always good to know that content resonates with the audience. The one takeaway I would say is the fact that AI needs to live on a spectrum. There needs to be an autonomy slider where you can set the “level of AI“ according to a human/AI collaboration spectrum.

Monique van Dusseldorp 
Kore.ai

OpenAI Customer Service AI Agent

A While back OpenAI dropped a GitHub repo with a slick demo of an AI-powered Customer Service Agent interface.

https://lnkd.in/dTWHfdZ6

Salesforce 𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 (Weiran Yao) released an 𝘦𝘯𝘵𝘦𝘳𝘱𝘳𝘪𝘴𝘦 𝘥𝘦𝘦𝘱 𝘳𝘦𝘴𝘦𝘢𝘳𝘤𝘩 (EDR) which is a multi-agent enterprise-grade deep research framework.

It consists of a feature rich GUI, and an underlying agentic workflow. With a master planning agent, and multiple worker sub-agents. Human-In-The-Loop, visualisation, reflection and MCP integration.

There are  two break-out use-cases for AI in the commercial world:

- code completion and
- deep research.

Both excel in personal setups and enterprise environments alike.

And they’re the break-out use-cases for two straightforward reasons…

First off, they live in environments with massive, high-quality contextual reference data.

Give AI abundant, relevant contextual data, and it generates succinct and accurate responses.

Sure, you could prompt Grok to create a white paper on the upsides of XYZ tech — but without that depth contextual reference, you’ll likely end up with generic fluff, the kind of “work slop” that is not useful.

Deep Research Agents — autonomous LLM-based systems that emulate human analysts by iteratively retrieving evidence, decomposing problems and constructing higher-order insights.

Second, success hinges on human oversight.

Vast amounts of research back this…

Agentic AI thrives on a spectrum, where you dial autonomy up or down based on the task.

Full autonomy for straight forward work, tight human reins for more complex and critical tasks.

It’s not about all or nothing automation — it’s about smart collaboration.

I would also add that inference latency is a significant impediment for Conversational UIs based on LLMs.

For obvious reasons deep research is not subject to the demands of latency. Which helps the use-case as-well.

Considering that the finding of data, vetting and synthesising take a vast amount of time, this is actually a use-case where inference latency pales against human driven effort time scales.

The paper, titled "Real Deep Research for AI, Robotics and Beyond" introduces and applies a systematic framework called Real Deep Research (RDR) designed to analyse large volumes of research papers. 

The aim is to uncover emerging trends, interdisciplinary connections and potential starting points for new inquiries. 

It uses  tools  or data filtering and reasoning, embeddings for projection and clustering, and expert-defined perspectives for structured overviews 

Focusing primarily on foundation models and robotics advancements, while briefly extending to other sciences. 

Enabling insights into what's currently happening...

New research from Microsoft, Agentic Organisations...scaling Agents...

What’s an Agentic Organization?

Imagine an entire company (or department) structured as a dynamic mesh of lightweight AI agents that:

- Split complex problems automatically
- Run multiple reasoning paths in parallel
- Collaborate concurrently
- Merge insights into a coherent final output

Microsoft Research just dropped a paper that opened my eyes wide on exactly this vision.

They introduce “AsyncThink” — a new reasoning paradigm where:

- An “Organiser” agent dynamically triages and forks sub-tasks
- Worker agents process in parallel (massively reducing latency)
- Results are merged with shared context

Their experiments? → 28% lower inference latency + higher accuracy on hard math reasoning, even compared to parallel thinking baselines.

https://lnkd.in/d3PcdgJP

I am facinated by the AI hype-cycle, and what is deemed as important and what not. For instance...

MCP captured the everyone’s imagination. But it is in ensense a protocol for vertical integration for AI Agents.

Recently two commerce  & payement protocols were launched (ACP & AP2) with little to no fanfare…

The same with model launches…the world stood still momentarlity when a new model was launched. Now it happens in stealth…

taplio

Cobus Greyling 's Linkedin Analytics

cobusgreyling

Performance Scorecobusgreyling's LinkedIn Performance

cobusgreyling's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Cobus Greyling 's Linkedin Analytics

cobusgreyling

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}cobusgreyling's LinkedIn Performance

cobusgreyling's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Cobus Greyling 's Linkedin Analytics

Performance Scorecobusgreyling's LinkedIn Performance

cobusgreyling's Best Posts (last 30 days)