Session library

How to Build AI Agents That Actually Do the Work

Gurleen Khurana · AIMG APAC · September 21, 2026

Gurleen Khurana explains how to build autonomous AI agents with persistent memory, server deployment, tool APIs, and model routing.

ai agentsautonomous ai agents

Gurleen Khurana explains how to build autonomous AI agents with persistent memory, server deployment, tool APIs, and model routing.

In this AI Marketers Guild session, Gurleen Khurana, Co-Founder and CEO of SimplifyGenAI, demonstrates how marketing and operations teams can transition from simple conversational chatbots to autonomous agentic workforces. Khurana shares the architecture behind live systems he runs for platforms like Cherry Shot, revealing how AI agents can independently conduct market research, publish optimized SEO articles, manage outbound sales pipelines, and interface directly with business operators over WhatsApp.

Rather than relying on manual prompting or advanced coding skills, Khurana illustrates how non-technical operators can direct autonomous agents through conversational harnesses, delegating complex technical setups like server management, Git-based memory deduplication, and tool integration directly to the AI.

Key moments

[02:23] What Is the Difference Between an AI Chatbot and an Autonomous AI Agent?

Gurleen Khurana explains that while standard AI chatbots require users to manually re-establish context and instructions during every interaction, autonomous AI agents operate with independent roles, persistent context, and long-term memory.

According to Khurana, an effective agent architecture rests on three foundational pillars: a defined operational role, continuous context retention that prevents the model from drifting, and memory partitioned into short-term operational buffers and long-term historical records. By solving the memory problem, agents eliminate the need for repetitive onboarding each time a task begins, enabling continuous workflows that persist across sessions.

Notable Quotes:

"the thing is that every time you have to give it a context you have to remind it hey this is what we are doing this is your role this is what you have to do." — Gurleen Khurana [02:01]

"The difference between a regular [Claude] or an AI agent is three things mainly. One is the role. What is that particular role of that agent is? Second is the context." — Gurleen Khurana [02:27]

"And the third thing is memory. The memory comes long form memory and short form memory. So these are the main components." — Gurleen Khurana [02:50]

[03:06] Why Run AI Agents Through a Terminal Harness Instead of Desktop Apps?

Gurleen Khurana explains that running an agent harness such as Claude Code directly in a command-line terminal grants the AI broad system-level access to local directories, files, and browsers, which standard desktop application wrappers restrict.

A harness serves as the operating environment through which an agent interacts with software and hardware. While standard desktop interfaces keep the model sandboxed within a closed chat interface, a terminal harness allows the AI to monitor downloads folders, inspect system files, trigger local scripts, and control web browsers to execute end-to-end tasks directly on the computer.

Notable Quotes:

"a harness is a way to use an AI agent. Currently if you're using [Claude] on your desktop you are using the [Claude] desktop as that's your harness" — Gurleen Khurana [03:15]

"what I use is I use [Claude] Code on my terminal because it gives AI full access to my laptop. So it can know what is happening in my download folders. It can control my browser. It can everything." — Gurleen Khurana [03:27]

[04:17] How Do Specialized AI Agents Communicate and Coordinate Tasks?

Gurleen Khurana describes an agent orchestration architecture where specialized agents communicate asynchronously through an internal messaging feed called "the hive" and execute tasks organized by operational priority.

In this multi-agent model, an orchestrator agent manages domain-specific sub-agents assigned to areas like search engine optimization, outbound email, and infrastructure monitoring. The agents publish task updates, blockers, and completed deliverables to the shared hive log. This shared context allows any agent or human supervisor to query the system and immediately understand what work has shipped, what is in progress, and which critical priority (P0 or P1) dependencies require attention.

Notable Quotes:

"hive is like the messages that they are sending to each other. They're sending messages to each other like what they needing to do what they are doing what they have done" — Gurleen Khurana [04:22]

"this way like whenever I talk with an agent it already knows what we are working on. I don't have to give it hey we are working on this, we are working on this." — Gurleen Khurana [04:52]

[06:22] Why Should AI Agents Be Deployed on Cloud Servers Rather Than Local Laptops?

Gurleen Khurana emphasizes that true autonomous agent workflows must be deployed on dedicated cloud servers so they can run around the clock on scheduled triggers without requiring a physical laptop to stay open and awake.

While a local computer can act as a server, deploying agents to cloud virtual machines allows them to run scheduled cron cycles independently. Khurana illustrates this with an agent that wakes up every morning at 9:30 AM, reads overnight communications from team members and other agents, pulls fresh performance data, and executes its assigned operational tasks. By integrating these cloud-hosted agents with channels like WhatsApp, human team members can query statuses, review reports, and issue directives from their phones at any time.

Notable Quotes:

"So the difference between running this on your [Claude] versus this system is that this is deployed on a server. A server is anything that has a RAM and a CPU. Your laptop is a server." — Gurleen Khurana [06:22]

"The only difference is here like I don't have to keep my laptop on. It is like in the cloud. So it wakes up. So every 9:30 it wakes up and then it knows what task it need to do." — Gurleen Khurana [06:37]

"It read the messages that we have sent it because we are like connected with it on WhatsApp. So every time we message it on WhatsApp is able to get on those messages." — Gurleen Khurana [06:55]

[08:38] How Can Non-Technical Operators Build and Orchestrate AI Agents Using Voice?

Gurleen Khurana explains that non-technical operators do not need coding skills to build sophisticated agent systems; instead, they can communicate their goals using voice-to-text tools and let an orchestrator agent write prompts and technical specifications for subordinate agents.

Khurana notes that using voice applications like Whisper Flow allows founders to convey complex reasoning and nuance far faster than typing. Rather than manually writing rigid prompt templates for every task, the operator speaks directly to a primary coordinator agent (such as "Trajan"), which then generates tailored system prompts and spawns focused sub-agents to research competitor strategies across Reddit, X (formerly Twitter), GitHub, and the broader web.

Notable Quotes:

"Since I'm not a technical person I don't know how to read or write a code but what I'm pretty good at is how to communicate with AI." — Gurleen Khurana [08:39]

"where you can like talk with the AI rather than like type it out because I feel like when you are talking you're able to express your thoughts and feeling in a better way." — Gurleen Khurana [08:50]

"I believe prompting is still like you know now where your school now you have the AI write the prompt for the other AI." — Gurleen Khurana [11:43]

[12:30] How Can Autonomous Agents Automate End-to-End SEO and Content Publishing?

Gurleen Khurana and Dax demonstrate how an autonomous SEO agent named Cherry Loop manages entire publishing workflows for the Cherry Shot platform, from competitor analysis to live deployment and backlink acquisition.

The agent queries SEO databases and Google Search Console to identify keyword opportunities, analyzes competing products (such as PhotoRoom) to gather pricing and features, sources visual assets directly from the server, and writes complete comparison articles. After deploying the content live, the agent submits the new URLs to Google Search Console to verify indexing and autonomously discovers software directories, submitting listing forms to secure external platform badges and backlinks.

Notable Quotes:

"Can you show how for example you know these agents are doing their job for example on cherry [shot] website how the agents are actually [shipping] pages how we also communicate with them on WhatsApp" — Dax [12:37]

"This is one of the blog post that it wrote comparing like our product with like photo room. It got the pricing off there by itself. It has access to the pictures and everything." — Gurleen Khurana [13:07]

"Recently we gave it like a video and be like hey can you like the video was someone using an CRM system like an open source CRM system. it watched the video and it created the CRM system." — Gurleen Khurana [15:15]

[17:33] How Do AI Agents Connect to Third-Party Software and Cloud Infrastructure via APIs?

Gurleen Khurana explains that AI agents integrate with external software, marketing platforms, and cloud infrastructure through standard Application Programming Interfaces (APIs), requiring users only to generate and provide API keys.

Rather than requiring manual coding, almost all modern software includes an API key generation setting. Once an operator pastes that key into the agent's environment, the agent can autonomously query endpoints, pull performance metrics, or control services. Khurana points out that he deployed and configured an entire Linux server on Hetzner without any personal systems knowledge simply by feeding an API key to his agent, which provisioned the machine and actively managed CPU and RAM usage to prevent system crashes.

Notable Quotes:

"API key and you can just like give that API key to the AI agent and the AI agent then will have access to that software. So after that like you don't have to be do anything technical" — Gurleen Khurana [18:13]

"The AI tell me hey like just go to this like [Hetzner] website sign in and then give me the API key and then after that it set up the server itself and then it's been running it ever since." — Gurleen Khurana [18:44]

[22:40] How Does Model Routing Optimize AI Agent Performance and Token Costs?

Gurleen Khurana explains that implementing model routing prevents agent systems from exceeding API budgets by directing specific tasks to the most cost-effective and capable large language model for that exact job.

Khurana structures his setup so that frontier reasoning models handle master orchestration, intent recognition, and planning. When secondary workloads are assigned, the system routes coding tasks to Claude Opus, broad web research to Claude Sonnet, high-volume processing to Meta's open models, and direct computer or graphical manipulation (such as Blender and After Effects) to Astra. This tiered structure ensures that expensive reasoning tokens are reserved exclusively for strategic decision-making.

Notable Quotes:

"what I have done is like I have given my AI instructions that the orchestrator the main thinker is going to be fable but when it's doing a research or when it's doing some coding then it uses other models." — Gurleen Khurana [23:01]

"when it does coding it uses opus when it uses research then it uses sonnet. So I don't like go over and waste all my tokens." — Gurleen Khurana [23:18]

"I believe Astra is pretty good at computer use. So in the sense of computer users like using Blender, working with anything 3D using After Effects and also just like controlling your computer in general" — Gurleen Khurana [24:29]

[27:09] How Does a Dedicated Memory Agent Prevent Knowledge Conflicts in Multi-Agent Systems?

Gurleen Khurana introduces "Dreamer," a specialized background agent responsible for nightly memory curation, deduplication, and consolidation across all concurrent agent operations.

Because multiple parallel agents continuously generate notes, client interactions, and project data, uncurated memory files can cause cross-agent conflicts and hallucinated discrepancies. Dreamer runs a nightly consolidation shift using Sonnet to pull recent activity records, remove redundant information, resolve conflicting operational notes, and commit clean, searchable summaries into a version-controlled Git repository. This process ensures that the collective organization maintains a single, reliable source of historical truth.

Notable Quotes:

"I'm Dreamer. I work the night shift curating our long-term memory, pulling the repo each day, using sonnet to consolidate knowledge and then committing it all to git. It's like tending a garden of memories." — Dreamer [27:27]

"Every night I gather all the important notes from the day, make sure they fit together smoothly by removing duplicates, and then store them permanently within us. It's like writing the day story down so nothing essential is lost." — Dreamer [27:48]

"because we are running so many parallel agents, right? You don't want memory of one agent to conflict with the memory of other agent. So it makes sure because here one part about the agents is like memory management." — Gurleen Khurana [28:33]

[29:08] What Security Safeguards Should You Implement When Granting AI Agents System Access?

Gurleen Khurana cautions that while agents require broad computer and software access to be effective, operators must establish strict security boundaries by withholding direct financial permissions.

To mitigate financial and operational risk, Khurana strongly advises against connecting AI agents to bank accounts, unmetered credit cards, or platforms where autonomous actions could incur runaway expenses. Restricting agent permissions to productive software tools, sandboxed environments, and specific task scopes allows teams to exploit autonomous execution without exposing their business to catastrophic monetary loss or accidental data corruption.

Notable Quotes:

"you might want to give only access to it the things that you feel comfortable with. I don't wanted to give access to my bank. I don't want to give access to anything that is related with money" — Gurleen Khurana [32:33]

"So I would like avoid that where you know like you have your credit card attached or where you know like it can do some stuff and you can lose a lot of money. So I wouldn't give access to that." — Gurleen Khurana [32:50]

[34:20] How Can Autonomous AI Agents Learn New Workflows and Self-Improve from Online Resources?

Gurleen Khurana highlights that operators can continuously enhance their agents by instructing them to inspect open-source GitHub repositories, read online articles, and analyze video tutorials to upgrade their own toolkits.

Rather than spending hours learning software setups himself, Khurana directs his agents to study emerging frameworks (such as OpenClaw and Hermes), extract the most secure and lightweight components, and incorporate those features into their own codebase. When an agent initially claims it cannot perform an unfamiliar task, Khurana notes that challenging the model's self-imposed limitations and prompting it to think outside the box frequently leads it to find creative, successful execution paths on its own.

Notable Quotes:

"I found this repo which let agent research on Twitter OnX. I asked [Claude] hey like can you install this repo? I just gave the repo, installed it." — Gurleen Khurana [34:40]

"I don't watch YouTube videos. I don't watch anything. I just ask it hey can you watch this video for me and get the information and then apply the information to yourself and then I do this every week" — Gurleen Khurana [35:19]

"AI will be like hey you know it will just speak out of memory or what hey I cannot do this and you just have to remind you hey you are a AI you are pretty smart and you can do it" — Gurleen Khurana [36:26]