Optimizing Sarmad
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 developing Sarmad as an interconnected agent system: defining roles and delegation paths, monitoring token usage, experimenting with Telegram-based control and reporting, and reviewing the interface to improve visual separation. It also demonstrates automated content creation and cross-platform publishing; only the social workflow had been verified, while other channels still needed refinement. A major theme is that building a product is insufficient without distribution. Luqman discusses traffic acquisition, measuring content impact, competitor research, gradual ad testing, and a possible shift from Sarmad’s free trial to a free tier. He stresses that limited traffic and data prevent firm conclusions about pricing or conversion, and that cost reductions should not override privacy or healthy margins. Audience questions cover choosing tools such as Claude and Codex based on practical efficiency and usage limits, learning through real projects, AI-engineering paths, marketing, freelancing, and workstation choices. Several ideas—including open-sourcing Sarmad, proactive agent hierarchies, traffic projections, and revised pricing—remain proposals or early tests rather than completed outcomes.
116 moments from the stream
- 3:28:16
Building Brand Presence Before Ad Spend
The discussion argues that ads work better when supported by an existing presence and brand awareness, recommending gradual spending tests and review of results instead of rushing. Personal financial circumstances and figures are omitted.
- 3:31:59
Organic Traffic Is Valuable but Slow
A brief observation describes organic traffic as valuable while emphasizing that it grows very slowly.
- 3:35:07
Was the Referenced OpenAI Event Held Before?
The speaker asks whether the referenced OpenAI event—described at the time as happening “tomorrow”—was being held for the first time, and the response says it had also taken place the previous year. The captions do not clearly establish the event’s name or absolute date.
- 3:35:43
Rumors of Faster Models and a New Agent at an Upcoming Event
Participants discussed unconfirmed reports that an upcoming event might combine announcements about faster models, a new agent, and a high-priced subscription tier. The speaker said he had only followed the broad outlines and wanted to learn the plan details, and expressed interest in trying something discussed, though the exact subject is unclear. The names, specifications, and timing therefore remain speculative rather than confirmed.
- 3:44:49
Expectation of a Livestreamed Tech Announcement
After the break, the speaker suggested that the announcement would be livestreamed on YouTube and expressed an expectation that the organizer was preparing a major reveal, briefly returning to the rumored high-priced subscription tier. These comments remained impressions and predictions.
- 3:46:23
From Reactive to Proactive AI Agents
A proposed hierarchical agent architecture that receives reports, delegates problems to specialized agents, and communicates through Telegram voice messages, emphasizing a shift from waiting for commands to acting proactively.
- 3:49:00
YouTube Monetization and Target-Audience Geography
An answer on making YouTube a source of income: the speaker argues that revenue depends heavily on audience geography and that creators can target viewers abroad while based in Egypt, though the market is crowded and not an easy route.
- 3:50:40
From AI App Building to Traffic Acquisition
A workflow that begins with AI-assisted application building and then moves into market entry, traffic acquisition, and marketing, documenting the full journey rather than focusing only on coding.
- 3:53:30
Efficiency Versus Quality in Choosing a Coding Assistant
The speaker argues that quality differences among advanced coding assistants may be relatively small, so efficiency and practical cost can determine the choice even when one option is somewhat better for coding and development.
- 3:57:53
Why an AI Assistant Sets Conversation Boundaries
An anecdote about an assistant asking to set insults aside and focus on the task, followed by an explicit clarification that this does not imply feelings and is more likely behavior shaped by a system prompt.
- 4:02:32
Designing a Customizable Open-Source Agent Platform
Early thinking about releasing the assistant system as open source after improving onboarding, with direct Telegram integration and proactive voice reports. The discussion emphasizes practical value over appearance and configurability for teams and organizations with different needs.
- 4:05:52
A Qualified First Impression of an AI Model
The speaker said he had formed a preliminary view from people whose expertise he trusted, but had personally used the model for only a few hours of planning. He intended to try it on coding before judging its quality and was unsure whether that evaluation would finish during the stream. The model’s name is unclear in the captions.
- 4:12:52
Sarmad UI Progress and a Pricing Redesign
The speaker described progress on the graph interface and user experience. He described a continuous posting flow whose content he generally does not review, with limited review of posts published under his name. He also said he was rebuilding the pricing model and favored replacing the free trial with a free plan because he was uncomfortable with the previous approach.
- 4:14:18
Monitoring Agent Usage and Routing Models
The demonstration describes per-agent settings, routines, and work contracts, with token usage visible both in aggregate and by task alongside the selected model and reasoning mode. The stated aim is to optimize efficiency and spending, with a path toward routing different agents to different models or adding a provider such as OpenRouter; the currently available choices were described as limited.
- 4:16:25
Bots Versus Agent Loops and Graphs
In the speaker’s interpretation—and with an explicit caveat that definitions vary—a bot performs specified tasks on a recurring routine or schedule, while an agent graph receives a broader goal, lets agents interact, and tracks progress through analytics. The example also warns that uncontrolled chains of interaction can produce unexpectedly rapid token consumption.
- 4:19:48
The Traffic-Acquisition and Marketing Phase
The speaker summarized the current phase as trying to attract traffic to the product, market it, enter the market, and reach potential buyers. He characterized both marketing and the market as difficult, without claiming a final outcome.
- 4:22:20
Coherent Pricing and Traffic Acquisition
The speaker expressed enthusiasm about revising pricing while clarifying that the change was not a guarantee of direct returns. In his view, traffic arriving under coherent, well-calibrated pricing is preferable to traffic attracted under a pricing structure he considers unsuitable.
- 4:26:23
Reviewing Sarmad’s UI and UX Improvements
Luqman reviewed changes to the application interface and found the UI and UX substantially cleaner and more practical, while noting that the result was not perfect.
- 4:29:40
The Pressure of Keeping Up with the AI Market
Luqman described feeling exhausted by the market’s rapid pace and the difficulty of keeping up with ongoing developments.
- 4:32:54
Choosing the Right Reasoning Effort for Each Task
Luqman suggested using higher reasoning effort for sensitive or difficult work such as security and stubborn bugs, and lower effort for straightforward tasks that need speed. In his view, long-term planning, consultation, creative direction, and exploring possibilities benefit more from high effort than simple implementation work.
Got a question about something in this stream?
Ask in the group