How to Ship Side Projects From Your Phone With AI Agents
For Indie developers and side-project builders · Based on Solmaz On-Demand Disposable Agent Orchestration Framework
// TL;DR
If you want to work on multiple side projects while commuting or away from your desk, the Solmaz framework's parallel channel workloads and Telegram Driven Development let you do it from a phone. Bind each project to a dedicated chat channel — Project-A, Project-B, Project-C — via ACPX and ACP, where each channel becomes a full IDE session driven by an on-demand disposable agent. You dispatch, monitor, and iterate across 1–5 channels simultaneously as a single operator. Use extra channels to retrieve artefacts agents can't send directly, like generated PDFs placed in /tmp. It compresses your elapsed build time dramatically without a desktop IDE.
How do you code without a desktop IDE?
You bind a chat channel to a harness session via ACPX. Each Discord, Slack, or Telegram channel becomes effectively a full IDE session driven by an on-demand disposable agent — a real Kubernetes pod with a full compute environment behind it. This is Telegram Driven Development (TDD): dispatching, monitoring, and iterating agent tasks entirely through messaging platform channels instead of a traditional editor. From your phone on a train, you're driving a full development environment.
How do you work on several projects at once?
Use parallel channel workloads. Instead of one agent per session, run one agent per task channel — name them by project or by harness plus index, like Project-A, Project-B, Project-C, or Codex-1 through Codex-5. Each channel is an independent full session. As a single operator you monitor across 1–5 channels simultaneously, dispatching work to one while another runs. This compresses elapsed time dramatically: your projects progress in parallel instead of serially waiting on your attention.
How do you retrieve files the agent can't send?
Agents can't always push arbitrary artefacts back through a chat platform — a generated PDF, a build output, a binary. Use a dedicated channel to retrieve them: have the agent place the artefact at a known path like /tmp, then pull it through a channel wired for that purpose. This closes the loop so TDD isn't limited to text-only exchanges and you can actually ship real deliverables from your phone.
What should you hand to the agent versus keep yourself?
Apply agents generously — treat the capability like an ointment for any problem agents can solve, with the default posture of "how do I take myself out of this loop?" But respect the boundary between shallow and fundamental work. Let the agent run shallow bug loops — uncovering and fixing superficial issues iteratively. Keep fundamental design decisions and architectural refactors for yourself; looping an agent on those produces slop. Your creative direction stays with you; the mechanical scaffolding goes to the agent.
How do you avoid the telephone game?
Don't relay instructions through a middle-model — asking one LLM to paraphrase your intent to another introduces errors because exact wording matters. Route directly through ACP so your instructions reach the target harness intact. If you catch yourself manually clicking or paraphrasing between agents, treat that friction as a first-class signal: it's the next thing to automate.
The result is a mobile-first, multi-project build workflow. You're not tethered to a desktop, you're not blocked on any single project, and each channel gives you a full disposable compute environment on demand.
Next step: Pick two side projects, stand up ACPX bound to a harness like Codex or Claude Code, create one Discord channel per project via ACP, and dispatch a first shallow task to each — then add a retrieval channel wired to /tmp so you can pull build artefacts back.
// FREQUENTLY ASKED QUESTIONS
Can I really run a full IDE from a chat app?
Yes — each channel bound via ACPX to a harness session is backed by an on-demand disposable agent running in a full Kubernetes pod, not a constrained sandbox. You dispatch and iterate through messages, and the agent operates a real compute environment. This is Telegram Driven Development: the chat channel is your interface, but a full IDE-equivalent session runs behind it.
How many projects can I run in parallel?
Typically 1–5 channels simultaneously as a single operator, using parallel channel workloads. Each project gets its own dedicated channel and its own disposable agent session, so they progress concurrently rather than waiting serially on your attention. Beyond about five, monitoring overhead starts outweighing the parallelism benefit for a solo developer.
What if the agent starts producing low-quality code?
You're likely letting it make design decisions inside a loop, which produces slop. Keep the agent on shallow bug loops — superficial fixes it can uncover and resolve iteratively — and pull fundamental refactors or architectural choices back to yourself. The shallow-versus-fundamental distinction is your quality gate; agents are safe for scaffolding, not for creative design from scratch.