Meet the Swarm
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
A wide-ranging, practical stream about building AI products in public. Its strongest thread is the shift from short tasks to long-running “missions” pursued by agent swarms, supported by a control center and living roadmap that expose progress, quality, and blockers. Luqman discusses requirements documents and system design as foundations for greater agent autonomy, splitting work between Codex, Claude, or specialized models, and measuring real cost per completed task rather than advertised token prices. These are presented as concepts, early experiments, and personal judgments—not settled results. Practical activity includes reviewing an interface and identifying a scrolling bug, requesting an interactive mock to clarify the design direction, and beginning an application walkthrough by granting access to a sample project and requesting a quick capability test. The stream also discusses web testing and continuous monitoring, with the explicit caution that agent-driven UI checks do not prove scalability. Audience questions broaden the stream to RAG and learning resources, context compaction, usage limits and tool handoffs, handwriting OCR, reviewed data-entry automation, and disclosed AI assistants for social engagement. It also covers app frameworks and community platforms, product validation, free-plan pricing, distribution and sales, SEO, ads, direct outreach, freelancing, and uncertain effects of AI on jobs and SaaS. Open questions remain around Mission Control’s interface architecture, subagent concurrency, the value and cost of bridging models, and whether several proposed integrations and workflows will work in practice. The supported conclusion is that documentation quality, low-friction onboarding, human oversight, realistic usage measurement, and distribution matter at least as much as model capability; many demonstrated products and workflows remained exploratory or unfinished.
124 moments from the stream
- 2:52:17
Specialized Agents and Pull-Request Monitoring
The speaker explains that subagents may specialize in tasks such as security, though he sometimes finds their behavior disorganized. He contrasts this with a separate proposed “watcher” that a user could enable to monitor GitHub pull requests continuously.
- 2:55:40
How Changing Usage Limits Affect Coding-Tool Choices
The speaker discusses how an unchanged price may still offer less value when usage allowances shrink, and how changing limits can prompt a reassessment of how work is distributed across tools. Some allowance estimates are uncertain and are not presented as definitive facts.
- 2:59:29
Preserving Skills and Context with a Coding Assistant
An audience suggestion recommends reusable skills and context files, while the speaker says he prefers working with two tools because their files and context have become easier to overlap in his experience.
- 3:00:35
Planning a Working Fork and the Application Roadmap
The speaker frames two next-step questions: whether to create a working fork before later developing the public application, and how to add a section for tracking its roadmap and mission. Part of the closing wording is unclear in the captions.
- 3:02:14
Balancing Context Continuity with Multiple Model Perspectives
One view argues that keeping a task with one assistant reduces the cost of rebuilding context. The speaker counters that a bridge between tools may mitigate this and that occasionally consulting a second assistant can provide an independent view of the project, while acknowledging that such a bridge takes time and the context problem is real.
- 3:05:24
Splitting Work Between Codex and Claude with Shared Project Context
The speaker considers using Codex for planning and prompt preparation before handing implementation to Claude, while stressing that the value of this workflow depends on strong documentation and system design. A new or compacted session must rebuild context from summaries and project files; when both tools can access the same repository and documentation, they can begin with broadly comparable project context, though understanding may become harder as the project grows.
- 3:10:49
Bridging Two Models with a Knowledge or Context Graph
The discussion suggests bridging two models through a knowledge graph or context graph. In the speaker's view, combining two strong models can open more options, but it may also consume additional tokens, making the approach potentially useful while requiring careful control of its costs and tradeoffs.
- 3:12:00
Optimizing Agent Usage by Task Type
Optimization starts by measuring each task's usage and recording which model handled it, then assigning model capability and reasoning level according to the work. The speaker recommends, as a practical opinion rather than a fixed rule, using stronger reasoning for managerial or coordinating roles and lighter settings for suitable execution tasks to reduce consumption.
- 3:19:02
Cost per Task Matters More Than Advertised Token Price
The speaker argues that price per million tokens alone can be misleading: output length and the total tokens needed to finish a task may make an apparently cheaper model more expensive in practice. He stresses evaluating this through actual usage rather than relying only on claims that a model is cheaper.
- 3:21:35
Model Routing and Harness Engineering
The group discusses routing small, focused tasks to lighter models while retaining a stronger model for orchestration, aiming to reduce token use and potentially improve quality. The speaker cautions that imposed rules or system-prompt changes can alter system behavior and long-term output quality, making harness engineering a useful but double-edged technique. He also says, based on his own observation, that one coding tool already routes some work across models automatically, though the tool’s name is unclear in the captions.
- 3:26:26
Orchestrating a Multi-Agent Workflow and Routing Tasks
The speakers discussed a role-based workflow that routes research, exploration, implementation, and advisory tasks to different agents or models. They presented it as a possible way to improve quality, speed, and cost efficiency, while warning that excessive model constraints could damage results and noting that quality and speed may matter more than minimizing cost. Any stated improvement figure was an estimate, not a documented test result.
- 3:29:00
A Claude Learning Resource
The speakers identified a public course said to be provided by Claude itself, although its exact title remained uncertain in the discussion.
- 3:33:51
The Need for Arabic Learning Resources
The discussion noted recurring interest in courses and observed that the referenced resource did not appear to be available in Arabic.
- 3:37:37
Sketching a Mission Control for Autonomous Agents
An early concept for a mission-driven control system emphasizes the importance of a broad map, considers—but does not commit to—a Kanban-like organization, and anticipates autonomous builders, agent swarms, and orchestrators. The discussion remains exploratory rather than a finalized design.
- 3:41:36
Automating Social Media Engagement Monitoring
The speaker outlines an agent that monitors and responds to comments and direct messages across social platforms to support community engagement and avoid missed opportunities. The workflow had been considered previously and was being revisited, with no evidence here that implementation or testing was completed.
- 3:50:41
A Brief View on Entering Construction
Construction was described as a good field with substantial sensitive paperwork, alongside the opinion that entering it alone may be difficult and that some backing or support is important.
- 3:53:18
Identifying a Scrolling Bug in a Review Interface
During an interface review, the left-side scrolling behavior and an incorrect horizontal bar were identified as broken, followed by a direct request to fix the behavior.
- 3:55:21
Using a Quick Interactive Mock to Test Design Direction
An interactive mock was requested so the design direction could be experienced in practice. It did not need to be perfect; the goal was to clarify the intended direction before focusing on the presentation deck.
- 3:58:01
From Static Interior Designs to AI-Assisted Interactive Walkthroughs
A discussion of turning interior-design images into an experience users can walk through, including video or VR viewing. It also proposes an integration layer between AI and engineering software such as AutoCAD to produce an initial plan draft, while explicitly expecting many early errors and retaining the engineer’s role in review and correction.
- 3:59:38
Validating a Product Idea Through Competitors and Domain Experience
Preliminary guidance for validating a new idea by finding competing products, studying their market position, and examining how they implement the concept. The speaker argues that firsthand domain experience can reveal real problems worth solving, with AI potentially assisting during implementation.
Got a question about something in this stream?
Ask in the group