Social Done. Search Next.
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
The stream centers on experiments with a multi-agent system and an automated, multi-platform social-content pipeline. Luqman discusses a possible shift from fixed loops and assigned tasks toward graphs and goal-driven agents with roles, contracts, memory, and hierarchy. He presents these as evolving ideas and experiments, not proven capabilities: agent communication cost, usage limits, stop controls, review, and reliability remain open problems. The practical work includes plans to generate video, text, voice-over, and captions, run the pipeline overnight with scheduled checks, and move gradually from human review toward automation while keeping personal posts under the account owner’s review. The stream does not establish that this end-to-end run was completed successfully. Audience questions broaden the discussion to connecting WhatsApp with AI tools, selecting models by total task time, retries, and usage rather than names alone, and protecting human judgment from model flattery or excessive autonomy. Other substantial topics include validating user problems, pricing around measurable usage costs rather than outcomes that cannot be guaranteed, the difficulty of acquiring traffic, and learning marketing, programming, and practical skills in the AI era. Additional discussions cover app-publishing and incorporation considerations in general, non-definitive terms; fair and secure agent competitions; visual summaries for long texts and streams; a proposed experiment with AI-generated musical variations; and lighter sports and workspace topics. The clearest conclusion is a preference for hands-on learning and building with human oversight and explicit operating limits. GitHub releases, autonomous operation, the virtual-team architecture, and several proposed tests and integrations remained plans or unresolved questions rather than confirmed results.
82 moments from the stream
- 0:05:03
Collecting Daily Reports Through WhatsApp
A question asks whether WhatsApp Cloud can collect an employee’s daily reports and process them regularly. Luqman says the idea is possible while expressing his view that WhatsApp creates complications for AI integrations.
- 0:06:42
A View on Islam Makhachev Moving Up a Weight Class
Luqman shares his view that Islam Makhachev may have effectively closed out his current division and could move to the next weight class. It is presented as personal speculation in a sports discussion, not confirmed news.
- 0:07:29
Paths for Connecting WhatsApp to an AI Tool
Luqman suggests first checking whether WhatsApp offers an API or another supported integration, since that would be the easiest path. If none is available, he mentions computer-use automation as a possible alternative and gives a general example involving a one-time WhatsApp login, retrieving the daily report, processing it, and returning it in the chat.
- 0:09:19
Choosing Marketing Topics for a Telegram Group
Asked about adding a marketing section to a Telegram group, Luqman says the intention is to choose a limited set of important topics instead of creating dozens of sections. He views marketing as especially valuable and difficult, while making clear that the plan is not yet completed.
- 0:10:29
Social channels configured; content volume remains
The speaker says LinkedIn, X, Facebook, and Instagram have been configured, with content volume still to be decided.
- 0:11:47
Customizing multi-agent automation and usage controls
The speaker describes an automation system whose features can be added or removed, with per-agent usage tracking, model and reasoning controls, and agent-to-agent communication. In his view, these controls can reduce token use and improve efficiency.
- 0:12:47
Planned GitHub release and the earlier command center
The speaker says he intends to publish the system on GitHub after activating it first. He also recalls publishing an earlier command center in April or May during what he calls the “loop engineering” period.
- 0:13:13
Loop engineering versus graph engineering
The speaker contrasts loop engineering, where a fixed workflow repeats, with graph engineering, where agents receive a goal and may react, coordinate, and choose tasks based on performance. He presents the shift from assigning tasks to assigning goals as a future vision, not as a result demonstrated here.
- 0:15:26
An attributed view on moving agents from loops to graphs
After initially hesitating over the name, the speaker recalls Peter Steinberger and attributes to him, through OpenClaw, an influential role in popularizing the loop idea. He also cites a tweet from roughly six weeks before the recording that asked whether the field had moved from loops to graphs, saying it prompted interest in “graph engineering.” This is the speaker’s recollection and assessment, not an independently verified history.
- 0:19:39
Choosing an agent platform with Claude support
The speaker says the platform under discussion can use Claude and recommends it as an alternative to another tool or model whose name and version are unclear in the captions.
- 0:20:49
Turning a Clip Workflow into a Reusable Tool
A clip-preparation workspace allows playback-speed adjustments. The discussion explains that publishing it through GitHub would require separating it from one project's needs and designing it for varied purposes and projects.
- 0:23:16
Validating User Problems Through Direct and Indirect Research
In response to a question about building before asking users, the speaker acknowledges interviews and surveys as useful but says he relied on indirect research. He argues that some problems are already visible, while presenting problem size as a possible factor in the scale of the return.
- 0:24:48
AI and the Coming Education Challenge
The speaker predicts that education and teaching will face a major challenge from AI, citing effects already appearing in Western countries, and views solving it as a potentially large-scale opportunity. This is the speaker's forecast, not an established result.
- 0:25:31
Why Price a Product by Usage Rather Than Outcome?
Addressing a question about pricing by customer benefit, the speaker says current pricing follows the service inputs that consume the most cost. He favors committing to a measurable volume of work instead of an outcome that cannot be guaranteed quickly, while also studying competitors' pricing and costs; he notes that prices may change.
- 0:29:31
A Repository That Generates a Virtual Team Structure
The speaker describes a repository that could turn a description of a desired organization into a framework of virtual roles, including system prompts, contracts, and hierarchy. He says the immediate goal is fast internal use, so this remains an in-development aspiration rather than a completed capability.
- 0:31:07
Balancing Automated Publishing Volume, Quality, and Review
The speaker argues that publishing across multiple platforms may expand reach, but excessive scale can reduce content quality. He proposes a workflow that automatically publishes items left in review for a defined period, excludes personal accounts, and sends notifications as content moves through review.
- 0:32:32
What Is an AI Agent Harness?
The speaker and an audience member explore what an “agent harness” means, distinguishing it from simply letting a model access an account or repeat an action. Tentatively, he associates it with tools, loops, and feedback that make a model more autonomous, noting that execution quality may depend on the model, its operating environment, and browser integration. He recommends testing a capable model suited to the task while explicitly saying he had not personally tested one referenced computer task-teaching feature.
- 0:40:25
Usage Limits and Reliability in AI Assistants
An audience question prompted discussion of rapidly consumed usage limits and the observation that some assistants may stall or fail to respond during a task. The comparison was practical and usage-dependent, not a controlled test.
- 0:40:58
Automating Content Review While Protecting Personal Posts
A proposed workflow handles a high volume of posts by allowing automatic publishing when content is not reviewed within a set window, while exempting personal posts so the system does not write in the account owner's voice without review.
- 0:45:06
Evaluating Model Efficiency by Time, Iterations, and Usage
The speaker suggested choosing a model by the task's total completion cost: time, number of iterations, usage, and the chance of finishing in one attempt. A heavier model may still be preferable if it reduces work and retries, while the decision varies by task and coding may consume more capacity.
Got a question about something in this stream?
Ask in the group