AI dev team
Can a manual back-and-forth with one AI tool be replaced by a coordinated system of agents, with one control point in Telegram and a clear chain of review?
Problem
Shipping a digital product takes a chain of hats: task definition, architecture, frontend, backend, QA, fixes until merge. With a single AI tool, I'm the one carrying context between steps, and that's the bottleneck.
Flow
- Telegram IntakeA business requirement or dev task. Gate: clarifying questions if needed
- OrchestratorBreakdown, routing, agent selection
- Architect (on demand)Only for structure, DB or dependency changes
- FE / BE WorkersWrite code within the defined context, with logs, files, status
- Deterministic QABuild, tests, lint. Failure goes back to the orchestrator to re-plan
- QA LLMDoes the output meet the requirement? Reasoned PASS/FAIL
- Decision in TelegramMERGE / REJECT / HOLD. The human decides
- GitHub PR MergeChanges merge only after approval
A shared TeamState holds context, decisions and blockers. Agents don't chat freely with each other: information flows through the state.
How it evolved
| When | What happened | What it taught |
|---|---|---|
| Feb 2026 | Architecture research, orchestrator vs. tech lead, budget and token caps | Separate product governance from engineering |
| Apr 2026 | Python/Redis/LangGraph build with Product, TL, FE, BE, QA, DevOps, Architect agents | Agents and routing worked, no end-to-end proof yet |
| Apr 2026 | Sign-up page test failed: GitHub clone 403, empty workspaces, retries with no commit | Environment, observability and stop conditions are part of the product |
| Aug 2026 | StateGraph + TeamState, two-layer QA, loop back to orchestrator, Telegram approvals, GitHub PR | A mature, documented architecture |
The Telegram experience
A chat that only says “done” hides failures. So I defined structured events, each with a task_id and context, so my reply returns to the right agent instead of getting lost as a new message.
- Less noise: short stage summaries instead of every token
- More control: PAUSE / RESUME / APPROVE / REJECT
- Resource control: iteration and budget caps, loop stopping
- Audit trail: who decided, which file changed, what QA said
Build vs Buy
| Alternative | What it offers | How mine differs |
|---|---|---|
| Claude Code / Codex / Cursor | A strong coding environment with one agent | Multiple roles and external governance via Telegram |
| Devin / Jules | Autonomous development in various forms | Custom control over routing, QA and merge conditions |
| GitHub Copilot | Built into the developer's workflow | A whole team, not code completion |
| CrewAI / LangGraph | Orchestration frameworks | LangGraph was a building block, not a competitor |
The open question: do several agents really beat one strong coding agent on ROI? Without a rigorous comparison on quality, bugs, time, cost and intervention, I have no grounded claim of superiority, and I don't make one.