Blog

The One-Person Dev Team: How to Run a Codex + Claude Code Agent Swarm with Zaki

Stop managing AI coding agents manually. Learn how to use Zaki as your [orchestration layer](/blog/the-52-billion-enterprise-ai-agent-opportunity) — spawning Codex and Claude Code agents in parallel, monitoring PRs automatically, and shipping like a team of 10 while staying a team of one.

February 24, 2026•12 min read•By Alaa Sumrain
💡

Quick Answer (AEO): You can run a dev team with AI by using a two-tier system: Zaki holds the business strategy and context, while coding agents (Codex/Claude Code) handle the execution in isolated worktrees. This allows a solo developer to ship dozens of PRs daily without manual micromanagement.

94

commits in one day

zero editor opens

7 PRs

in 30 minutes

idea to production

~$190/mo

total AI spend

Claude + Codex + Zaki

That is not a demo. That is a real B2B SaaS founder who had three client calls the day they shipped 94 commits. They did not write a single line of code.

The architecture behind it runs on Zaki. And you can replicate it today.


The Problem Nobody Talks About When You Start Using Coding Agents

You discover Codex or Claude Code. You run it on a task. It works. You get addicted to the leverage and start throwing more at it.

Then reality hits.

💡

Context windows are zero-sum. Every token you spend explaining your business — who the customer is, what you tried before, why it failed last time — is a token not spent on code. Agents loaded with business context run out of room for the codebase. Agents loaded with the codebase have no idea what you're actually building or why.

You end up micromanaging. Writing prompts manually for each task. Babysitting three terminal windows. Losing track of which agent is on which branch. Getting a CI failure notification with no idea which task caused it.

The leverage disappears under the weight of overhead.

The fix is a two-tier system. That is exactly what Zaki gives you.


What is the "Two-Tier" architecture for AI dev teams?

The insight that changes everything: specialization through context, not through different models.

Tier 1 — Orchestrator

🐙 Zaki

Business context · Memory · Customer data · Decisions · SOPs

Scopes tasks · Writes prompts · Monitors progress · Notifies you

↓ spawns ↓

Tier 2 — Workers (isolated worktrees)

⚙️

Codex

Complex logic
Backend · Refactors

🤖

Claude Code

Fast frontend
Git ops · UI fixes

✨

Gemini

Design specs
Visual layouts

Each agent: own worktree · own branch · own tmux session · never touches prod DB

Zaki holds (business layer)

  • →Customer conversations and feature requests
  • →Meeting notes and what was promised to whom
  • →Past architectural decisions and why
  • →What failed before and the lessons from it
  • →Your SOPs, brand voice, prod DB access

Coding agents hold (execution layer)

  • →The codebase — and only the codebase
  • →The specific task and acceptance criteria
  • →Type definitions, test file paths, PR conventions
  • →No customer data, no prod DB, no business history
  • →Isolated worktree, isolated branch, isolated session

Zaki does not write code. It understands your business at the strategic level and translates that understanding into precise, contextual prompts for your coding agents. This is the core of the Humans for Ideas, Zaki for Execution model.

Your orchestrator stays sharp on strategy. Your workers stay sharp on code. Neither is doing the other's job.


What is the full workflow from request to merged PR?

1

Customer request → Zaki scopes it

A customer asks for a feature. You describe the request to Zaki on Telegram or Discord. Because Zaki has your meeting notes, your architecture context, and your previous decisions, zero background explanation is needed. Together you scope it. Zaki then automatically tops up customer credits, pulls their config from your read-only prod DB for context, and writes a detailed prompt for the coding agent.

2

Agent spawns in an isolated worktree

Each task gets its own git worktree and its own tmux session. Agents do not share branches, do not share node_modules, do not step on each other. The task is registered in a JSON file that Zaki reads during monitoring.

3

Cron monitors on autopilot every 10 minutes

A cron job checks the task registry deterministically — are tmux sessions alive? Do tracked branches have open PRs? Did CI pass? Are there critical review flags? It auto-respawns failed agents with an updated prompt. You are not watching terminals. The system tells you when to look.

4

Agent creates the PR (not done yet)

The agent commits, pushes, and opens a PR via gh pr create --fill. This does NOT trigger your notification. A PR alone is not done.

5

Three-model automated code review

Every PR is reviewed by Codex, Claude Code, and Gemini before you see it. Each catches different things. All three post comments directly on the PR. Agents with critical review failures get respawned with the feedback in their context.

6

CI runs the full test suite

Lint, TypeScript, unit tests, E2E, Playwright against a preview environment. Rule: if a PR touches any UI, it must include a screenshot in the PR description. Otherwise CI fails. This alone cuts review time in half.

7

You get the notification — one Telegram message

By the time you see it: CI is green, three reviewers approved, screenshots are attached, edge cases are documented. Your review takes 5-10 minutes. Many PRs get merged without reading the code.

8

Merge. Cleanup. Next task.

PR merges. Daily cron cleans up orphaned worktrees and archives the task entry. The system picks up the next item in the queue.

✅

Zaki's definition of done (all must be true before you get notified): PR created · No merge conflicts · CI green · Codex reviewer passed · Claude reviewer passed · Gemini reviewer passed · Screenshots included for UI changes.


How are AI agents spawned in this system?

# 1. Create an isolated environment for this task
git worktree add ../feat-custom-templates -b feat/custom-templates origin/main
cd ../feat-custom-templates && pnpm install

# 2. Spawn Codex in a dedicated tmux session
tmux new-session -d -s "codex-templates" \
  -c "/path/to/feat-custom-templates" \
  "codex --model gpt-5.3-codex \
  -c model_reasoning_effort=high \
  --dangerously-bypass-approvals-and-sandbox \
  'Your detailed, context-rich prompt here'"

# 3. Or spawn Claude Code for frontend work
tmux new-session -d -s "claude-ui-fix" \
  -c "/path/to/feat-ui-fix" \
  "claude --model claude-opus-4 \
  --dangerously-skip-permissions \
  -p 'Your prompt here'"

The task registry entry:

{
    "id": "feat-custom-templates",
    "tmuxSession": "codex-templates",
    "agent": "codex",
    "description": "Template system for agency customer — save and reuse configurations",
    "branch": "feat/custom-templates",
    "startedAt": 1740268800000,
    "status": "running",
    "notifyOnComplete": true
}

When complete, it updates with the result:

{
    "status": "done",
    "pr": 341,
    "completedAt": 1740275400000,
    "checks": {
        "prCreated": true,
        "ciPassed": true,
        "codexReviewPassed": true,
        "claudeReviewPassed": true,
        "geminiReviewPassed": true
    },
    "note": "All checks passed. Ready to merge."
}

What happens when AI agents fail?

Standard retry loops respawn a failed agent with the same prompt. That almost always produces the same failure.

⚠️

The Intelligent Retry Loop: If the prompt was insufficient the first time, respawning with the same prompt is not a retry — it's a loop. Zaki breaks the loop by rewriting the prompt using the failure reason + full business context.

What that looks like in practice:

  • Agent ran out of context window? → "Focus only on src/types/template.ts, src/api/templates.ts, and src/components/TemplateEditor.tsx. Ignore everything else."
  • Agent went the wrong direction? → "Stop. The customer wanted a save/load UI, not a config migration script. Here's their exact request from the call: [quote]."
  • Agent needs clarification? → "Here's the customer's email from last week explaining their workflow. Their team has 12 people sharing one account."

Over time, Zaki logs what works: "Codex needs type definitions upfront for billing features." "Always include the test file paths — the agent wastes 20 minutes finding them." The system gets faster with every task it completes.


How do I choose the right AI agent for the task?

⚙️

Codex

The workhorse — 90% of tasks

  • •Backend logic and complex bugs
  • •Multi-file refactors
  • •Reasoning across the codebase
  • •Billing, auth, data layer work

Use model_reasoning_effort=high for complex tasks. Slower but thorough. Low false positive rate on reviews.

🤖

Claude Code

Fast — frontend and git ops

  • •UI components and style changes
  • •Git operations and PR creation
  • •Copy updates and content changes
  • •Quick fixes under 5 files

Faster than Codex. Fewer permission issues. Review output tends toward overengineering — only act on critical flags.

✨

Gemini

Design sense — spec first

  • •New dashboard screens and layouts
  • •Visual design specifications
  • •Security and scalability reviews
  • •UI/UX feedback on existing pages

Workflow: Gemini generates HTML/CSS spec → Claude Code implements it in your component system. Better output than asking one agent to design and build.


How does Zaki find work without being asked?

This is where the system stops being a tool and starts being a genuine force multiplier. Zaki does not wait for you to assign tasks. It scans for work:

Every morning: Scans Sentry for new unresolved errors → spawns agents to investigate and fix.

After customer calls (triggered by calendar): Scans meeting notes for feature requests → spawns agents. Flags commitments made on the call → creates reminders.

Every evening: Scans git log → spawns Claude Code to update changelog and customer-facing docs. Checks CI failures from the day → creates bug tickets.

✅

The Outcome: You take a walk after a customer call. You come back to a Telegram message: "7 PRs ready for review. 3 features from today's call, 4 Sentry bugs fixed overnight." That is the system working the way it should.


The Economics

Claude API (via Anthropic)~$100/month
Codex API (via OpenAI)~$90/month
Zaki on Zakified (Starter)$49/month
Total~$240/month — less than 2 days of a junior developer

You can start even smaller — $20 in API credits and Zakified's starter plan proves the concept on real tasks.

The constraint is not budget. It is RAM. Each agent needs its own worktree and its own node_modules. 4-5 parallel agents on a 16GB machine is the practical ceiling before memory pressure starts slowing builds. If you are serious about running 7+ agents in parallel, invest in hardware: 32GB minimum, 64GB+ for heavy parallelism.


Setting This Up on Zakified

You do not need to build any of this from scratch.

1

Deploy your Zakified instance

Pick a plan and provision your server. Zaki is running in under 5 minutes. No SSH, no configuration files, no DevOps required.

2

Give Zaki your business context

Connect Telegram or Discord. Walk Zaki through your codebase, your customers, your architecture decisions. Zaki writes everything into persistent memory. It does not forget between sessions.

3

Connect your repo and set up the CLI tools

SSH into your server and install Codex CLI plus gh for GitHub. Zaki will walk you through the setup step by step if you give it your repo URL.

4

Set up the task registry

Create the .tentacledbot/ directory and a blank active-tasks.json. Tell Zaki your CI commands, test runner setup, and PR conventions — it includes this in every agent prompt automatically.

5

Run your first task

Tell Zaki about a bug or a feature. Watch it scope the work, write the prompt, spawn the agent. Check back in an hour. Your PR will be waiting.


What This Actually Means

The leverage here is not linear. It compounds.

Every task Zaki handles trains its prompting judgment. Every PR that ships teaches the system what works for your codebase. Every retry teaches it how to recover. Six months in, Zaki is not just an orchestrator. It is a senior engineer who has been with your company the whole time, knows every decision that was made, and never forgets anything.

💡

The one-person company of 2026 does not look like someone grinding 16-hour days. It looks like a founder with sharp strategic judgment, a Zaki instance holding full business context, and a fleet of coding agents executing while they sleep.

35 million US businesses need someone to build and operate systems like this for them. Founders who understand this architecture in 2026 are in the same position as people who understood cloud hosting in 2008.

The infrastructure exists. The agents are capable. The only missing piece is the orchestration layer that turns them from clever demos into actual business leverage.

That is what Zaki is.


Deploy Zaki and run your first agent task today →

Ready to deploy your Mission Control?

Stop managing servers. Start managing outcomes. Deploy your first Zaki agent in 60 seconds.

🐙

Alaa Sumrain

Building Zakified — dedicated AI agent hosting powered by OpenClaw. These guides come from running real agent swarms in production across dozens of business use cases.