How I Manage 45 AI Agents Across 3 Machines
DevOps & AI Toolkit · 2026-10-26
💡 Quick Take
1. The host created Agent Deck Desktop, an open-source app that monitors AI coding agents running across multiple computers in one view.
2. The app connects through secure shell (SSH) tunnels to native terminal tools like Claude Code and OpenCode without adding custom interface wrappers.
3. According to the host, developers should switch agents to auto mode once repository rules and test pipelines become reliable.
4. A dispatcher agent takes queues of 20 tasks and assigns them to isolated Git worktrees (linked working directories) across separate machines.
5. The host notes that automated dispatchers cannot yet talk across machines, so human managers must still assign separate task lists to each box.
📊 Detailed Explanation
The host explains how his workflow moved from micromanaging a single assistant to managing 45 agents at once. Early on, a developer approves every tiny terminal command, such as viewing a bug report. Once repository instructions and project skills become dependable, approving turns into a reflex that slows work down. According to the host, developers should switch mature setups to autonomous mode. In this pipeline, an agent reproduces bugs in an isolated Git worktree, writes code, and opens a pull request. Automated review bots like Qodo examine the patch across multiple rounds, and GitHub Actions agents verify all green checks before the host ever inspects the result.
Scaling up required dedicated hardware and orchestration tools rather than running everything on a personal laptop. The host built the open-source Agent Deck terminal app to coordinate multiple coding tools, including Claude Code, Codex, and OpenCode. A lead dispatcher agent pulls from an issue queue and assigns jobs to either single agents or specialized teams. For large feature documents, the orchestrator divides labor between a coder, reviewer, auditor, tester, and release agent. The host runs lighter workloads on an office mini PC with three parallel task slots. When local resources fill up, he spins up cloud servers to run six additional tasks simultaneously, keeping his daily workstation free of background noise.
Managing several remote boxes introduces a major visibility bottleneck. Jumping across separate terminal tabs makes it easy to lose track of finished jobs or agents waiting for input. The host built the desktop version of Agent Deck to surface every machine and agent status in a single window. Crucially, the app avoids wrapping coding agents in custom graphic overlays. The host advises against wrapper tools because underlying command-line agents update far too quickly. Instead, Agent Deck embeds the raw terminal session through background daemons. The host warns that the remaining bottleneck is human review: merging 18 pull requests at once causes cascade merge conflicts that invalidate prior approvals, while separate machine dispatchers cannot yet coordinate workloads with each other.
🎯 Tech Expert Opinion
The host correctly identifies that software engineering with AI has transformed into an infrastructure orchestration problem. Running 45 agents across three physical environments mirrors classical distributed computing patterns. Using Git worktrees to isolate file systems prevents agents from trampling each other during local edits. Furthermore, treating a primary model as a dispatcher that assigns tasks based on agent specialization creates genuine labor division. The data shows that generating dozens of pull requests in hours is solved; managing code integration is the true obstacle.
Preserving native terminal sessions over building bespoke visual wrappers is the most pragmatic technical choice highlighted in the video. Fast-moving command-line agents gain new flags, key bindings, and slash commands almost daily. History shows that proprietary user interface layers fall behind their underlying terminal engines within weeks. By relying on a lightweight background daemon communicating over standard secure network channels, the tool maintains full compatibility with native ecosystem updates while delivering fleet-wide monitoring.
A critical operational risk emerges once automated production reaches this velocity. As the host observed, 18 simultaneous pull requests trigger cascading merge conflicts. When one pull request merges into the main branch, Git invalidates the approval state of other pending branches. Without continuous automated rebase agents, human engineers drown in rebasing and regression tests. Furthermore, running dozens of agents simultaneously across cloud servers and desktop hardware introduces significant model execution bills and token consumption risks that require strict budget ceilings.
According to the host, this multi-machine orchestration pattern makes sense for senior engineers, maintainers with massive backlogs, and teams building continuous integration pipelines. Solo developers working on simple web apps or tightly coupled codebases should wait. Until dispatchers can communicate across networks to rebalance tasks and resolve integration conflicts automatically, running fleets of this scale requires advanced systems architecture skills that easily overwhelm casual programmers.
Kanal: DevOps & AI Toolkit