Four Projects. One Flow.
Want to back the stream? The «داعم» membership on YouTube is EGP 60 a month: badges, emoji, and I answer your questions first on stream. Become a supporter
The gist
Luqman and the audience explore a platform that would turn conversational or spoken goals into teams of agents for building products and running operations. Topics include roadmaps, role contracts, architectural oversight, and knowledge and context graphs; the concept remains under development and its feasibility unresolved. Token consumption, repeated file reading, review loops, and blocked parallel work recur throughout, alongside proposals for event-triggered supervision without a practical comparison establishing its advantage. Practical work includes reviewing an animated hoopoe website, exploring 3D and reverse-engineering tools, and demonstrating keyword search in a partially indexed stream library, with semantic search proposed. Content automation, reply approvals, and scheduling are also reviewed, including posting that reportedly continued after a pause request and unresolved performance and storage problems. Audience questions cover learning AI, choosing a specialization, and reaching customers; Luqman recommends planning distribution and using direct outreach for B2B. His predictions about voice interfaces and increasingly autonomous agents remain forecasts, while the work presented does not establish a completed platform or successful end-to-end automation.
153 moments from the stream
- 4:38:15
Token limits and AI worker settings
An audience question prompts a discussion of token limits. Luqman describes tracking usage per task and adjusting the model and reasoning level, and expects limits to become more important over time. This passage does not show limits being implemented.
- 4:40:43
From a user goal to teams that build and manage
While reviewing teaser wording, Luqman describes a concept where users state a goal conversationally and the system generates teams to build and manage it. He frames it for people with ideas and limited familiarity with AI coding tools. This describes intended behavior rather than demonstrating its implementation.
- 4:43:44
Limits of task-centered agent tools
Luqman discusses an unnamed tool’s tasks, routines, skills, connectors, and pull-request history. He feels its interface needs polish and warns that task-centered work without architectural oversight can produce tangled code. The discussion also considers different agent harnesses as a way to reduce cost, even when token usage increases.
- 4:45:46
Competition postponement and early-access plans
Answering an audience question, Luqman says the competition will be postponed and describes plans for early access followed by an open-source release. He links further competition planning to that release without setting a date.
- 4:46:16
Role contracts and team structure for an agent organization
The discussion outlines an organization with a mission, goals, and a contract defining each agent’s responsibilities, with interactions beyond task assignment. The example includes reporting relationships, engineering and QA roles, and an architect intended to help prevent tangled code.
- 4:48:51
Extending the hoopoe’s motion across page sections
Luqman requests improvements to the hoopoe character’s motion, describing missing animation between the hero and subsequent sections. He also proposes branches for the character to interact with. These are development requests rather than confirmed implementation results.
- 4:50:01
A mission-led, event-driven executive agent
The proposal gives the executive agent a mission and roadmap instead of relying solely on assigned tasks or continuous operation. Events, hooks, and triggers would wake it when needed to limit unnecessary token use; some event examples are unclear in the captions.
- 4:50:41
Architectural oversight and code inspection for broader agent context
The discussion considers the architect’s role in product development, inspecting the existing codebase, and producing information that supports decisions beyond a single session’s context. It also mentions microservices versus a monolith and organizing code into modules. Some workflow details are unclear in the captions.
- 4:52:09
Knowledge graphs for coordination and an introduction to context graphs
The discussion describes a knowledge graph that maps relationships and responsibilities: who owns each area, reporting lines, and assignments involving the database and RAG. It ties effective parallel work to architectural organization, then begins an unfinished context-graph example involving a production fix.
- 4:53:44
Connecting knowledge, event context, and agent harnesses
The discussion considers connecting relationships with events and actions to track what happened, where, when, and which agents participated, supporting shared workflow context. The speaker interprets the proposed architecture as combining graph engineering with agent harness engineering, while noting that the broader scope could require more development time.
- 4:57:34
A voice interface for building agent teams from goals
The speaker outlines a vision for a voice interface that receives a goal and assembles teams for execution and management. Answering an audience question, he distinguishes configuring agents from scratch from using agents with predefined work contracts and a configured harness. The proposal depends on substantial graph and harness engineering, which he cautions requires careful understanding.
- 5:07:52
Security question about a WordPress theme built with coding assistants
An audience member asks how to secure a WordPress theme built using Codex and Claude Code. Luqman acknowledges limited knowledge of WordPress security and provides no specific protective steps, leaving the practical question unresolved.
- 5:08:41
Choosing an Arabic name that fits a tool’s character
Luqman and the audience discuss choosing an Arabic name, weighing pronunciation, perceived strength, its connection to the tool’s interactive character, and the difficulty of finding a suitable domain. He expresses personal preferences without settling on a name.
- 5:13:20
Preconfigured AI teams selected for business needs
Asked whether team members are selected or generated from a goal, Luqman describes an idea still being explored: preconfigured roles that users choose according to their company’s needs, such as a sales lead. He begins explaining goal-based team generation, but the page ends before the explanation is complete.
- 5:21:21
Designing agent teams around user goals
A conceptual discussion of understanding a user's goal to identify skills, break down work, and assemble an agent team. A participant questions the difficulty of automatic agent creation at the outset; Luqman also proposes a manual option. No verified implementation is shown in the captions.
- 5:23:00
An unresolved Claude Code connector question
A question about adding connectors to Claude Code through a tool whose name is unclear. Luqman says he has not used it and does not know the solution, suggesting a search or asking the assistant itself.
- 5:23:23
Starting with AI from your background and goals
Luqman advises beginners to clarify whether they want to improve an existing skill or enter programming and software development. He suggests subscribing to Codex or Claude, explaining their background and ambitions, and asking for help reaching their goals.
- 5:27:28
Using AI to build and market products
Luqman describes using AI to build products and work on bringing them to market and selling them while sharing the process through livestreams. He points to the opening agenda for more context; no sales results are presented here.
- 5:30:47
Inviting community input on a product name
Luqman proposes sharing the website once it is finished and running a Telegram poll to choose among possible names. These pages do not show the poll being conducted or its outcome.
- 5:32:07
Documenting and sharing technical advice in Markdown
Luqman proposes saving technical advice in the repository and summarizing or converting it into a Markdown file for sharing through Telegram. He describes the format as easy for an AI assistant to read; the original advice is not provided here.
Got a question about something in this stream?
Ask in the group