Back to Warp.dev? 2x Claude + Codex
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 indexed portion of the recording centers on coding-agent orchestration. Luqman describes using Codex to supervise Claude sessions through computer use and scheduled checks, allocate tasks, hand off context, and review work. Audience discussions raise workflow visibility, lost details, and trust in reviews, particularly for scientific code; reliability and token efficiency remain unresolved. He explores a goal-driven workspace combining agent teams, memory, roadmaps, and an interface that adapts to each project. Practical work includes reviewing system architecture, extending plans beyond MVP, and requesting landing-page and hoopoe-animation revisions. The concept and revisions remain in development, alongside discussion of open sourcing; completed implementation and fixes for computer-performance problems are not established. Audience questions cover getting started with AI, programming fundamentals, hardware requirements, security, and developing skills. Other discussions examine task cost versus token pricing and model speed without a measured comparison. On product strategy, Luqman emphasizes lasting usefulness, discovering problems for a focused audience, and planning marketing and distribution. He argues that easier building does not guarantee sales and that marketing outcomes remain uncertain. Secondary topics include agents that read and reply to Facebook and TikTok comments while identifying themselves as AI. Luqman reports workflow problems, local slowness, and agents stopping when the application or computer closes, without documenting fixes. Participants also discuss voluntary creator support through tips and paid messages, affiliate marketing, sponsorships, personal branding, and professional visibility through building in public; the question of how personal branding relates to technical competence remains unresolved. Discussion also covers AI-company IPOs and profit-and-loss disclosure. At recording time, Luqman offers unverified forecasts about IPO timing and funding improving pricing and efficiency, while audience contributions suggest funding could support data centers and research. He also discusses his trust in expert forecasts and a warning attributed to Geoffrey Hinton about superhuman-AI risks, presented as a risk forecast rather than an established outcome.
160 moments from the stream
- 1:33:33
Choosing where the character enters the screen
Continues planning the envelope-carrying character's movement, considering entry from the top right or the right side. The entry position remains undecided.
- 1:35:28
A provisional onboarding preview and open-source plans
Describes the current work as an early onboarding preview whose final appearance will change. Notes many development milestones and the possibility of opening the source after establishing a strong foundation.
- 1:36:21
From suggested replies to intention-driven work
Discusses suggested replies in Claude and Codex as drafts users can review and change. Sees them as a direction toward agents executing goals with less manual prompting, presenting this as a project ambition while acknowledging uncertainty about achieving it.
- 1:40:25
Turning project documents into a trackable roadmap
Explains deriving an MVP and later-release roadmap from requirements and design documents, then organizing it into milestones and tasks. He argues that visual progress tracking matters when delegating implementation to AI.
- 1:44:27
Using a supervising agent to manage Claude sessions
In response to a question, describes giving a supervising agent access, starting with one trial session, and then delegating work allocation and follow-up. The workflow includes periodic checks of session output and files, plus messages between agents.
- 1:46:52
Discussion of Chinese model latency and Kimi for coding
Discusses a question about slow Chinese models and relays another participant’s favorable opinion of Kimi for coding. He suggests possible cost savings despite latency and considers exploring these models for agent harness engineering; the discussion establishes neither a specific version nor a performance comparison.
- 1:49:38
Refining the hoopoe’s entrance on screen
Finds the hoopoe’s current appearance too abrupt and requests trying a smaller size or an entrance from the right. The passage contains feedback and a revision request, without confirming an implemented result.
- 1:50:59
Automating handoffs with computer use and scheduled tasks
Explains that the demonstrated setup uses computer use and scheduled tasks, saying it does not rely on plugins. He describes automatically generated agent handoff prompts and points to one he did not write himself; checks mentioned in that prompt are not evidence that they ran.
- 1:57:50
An initial suspicion about a dictation problem
Mentions a problem with a dictation tool and suspects the CPU as a possible cause. The explanation ends unfinished, with no confirmed diagnosis, test, or fix.
- 1:58:15
Delegating periodic supervision of Claude sessions to Codex
Answering a setup question, Luqman describes asking Codex to access a Claude session and supervise a task or project milestone through scheduled checks, for example every fifteen minutes, reviewing progress and the codebase. He says he needed no special setup and personally prefers Codex’s computer use to Claude’s; these are his reported experience and opinion.
- 2:02:26
Automated handoff prompts and hard-to-follow agent exchanges
Luqman describes a session containing automated handoff prompts between agents. He remarks that replies seem different to him when one agent prompts another, making the exchange difficult for him to follow; this is a personal impression.
- 2:03:30
Orchestrating Claude and Codex sessions and handoffs
Luqman evaluates outputs through user experience and feels that assistant handoffs improve management of building and debugging. He describes an orchestrator managing sessions within one tool and values the different perspectives of Claude and Codex; these are impressions, not a demonstrated comparison.
- 2:07:56
Detail loss and trust in scientific code review
An audience discussion raises detail loss during Claude–Codex delegation in scientific coding. Luqman asks whether conflicting model answers explain the uncertainty about what to trust; no definitive verification method is established.
- 2:10:33
Building projects live with AI
Luqman describes livestreams about building tools and projects with AI, trying to bring them to market, and sharing the work and learning process publicly.
- 2:11:31
A coding and review loop across multiple assistants
An audience member describes trying a loop in which one assistant writes code, followed by successive reviews and a final check using other tools. Luqman says the workflow sounds solid, without demonstrating its effectiveness; the specific model names are unclear.
- 2:13:19
Marketing channels and an email campaign experiment
The discussion covers organic marketing, paid advertising, and an email campaign planned at the time of recording. Luqman considers the strategy straightforward but says its effectiveness remains to be tested.
- 2:14:38
Requesting better output quality and crispness
Luqman asks the assistant to improve output quality and notes that one element sometimes lacks crispness. The element’s name is unclear in the captions, and no resulting improvement is established.
- 2:15:33
Advice for students: using AI to develop skills
Luqman advises students to experiment and use AI for their studies and technical skill development, without treating immediate earnings as a requirement during learning. He also encourages ambition and a clear vision.
- 2:18:14
Learning programming fundamentals with AI
A reflection on learning programming fundamentals with AI, alongside appreciation for programmers’ expertise and what can be learned from them.
- 2:19:18
Expanding skills and entering new fields with AI
The speaker argues that AI can build on existing knowledge and help people explore other fields, including programming and marketing, while acknowledging that programming is more technical.
Got a question about something in this stream?
Ask in the group