amux for CTOs

Run an AI engineering team from one dashboard — self-healing parallel agents, shared kanban, mobile oversight, and a full audit trail. Updated July 2026.

The problem amux solves for engineering leaders

When an engineering team adopts AI coding agents at scale, three problems emerge immediately:

1. Fleet visibility: Dozens of Claude Code sessions running in tmux across developer machines — no shared view of what's running, what's done, what's stuck.
2. Coordination failure: Two agents claim the same task, overwrite each other's branches, or block on the same API rate limit while the other provider's quota sits unused.
3. Overnight reliability: Agents hit context limits at 2am, crash on tool-approval prompts, or silently hang — and no one knows until morning.

amux solves all three in a single self-hosted deployment your team controls.

Fleet overview at a glance

The amux web dashboard gives the entire team — developers, managers, and technical leaders — a shared real-time view of every active agent session. Status dots tell you at a glance: green = working, yellow = needs attention (YOLO prompt, approval required), red = crashed. The peek panel streams live terminal output from any session. You can watch an agent's work, read its plan strip (the current task list), and switch between its subagent transcripts without SSH.

From the iOS app on your phone, you have the same full view. Review fleet health in a standup, approve a risky action from the airport, or steer an agent mid-task via the steering queue — without opening a laptop.

How parallel agents coordinate without conflicts

amux's kanban board uses atomic SQLite compare-and-swap task claiming. The sequence is:

1. A developer or orchestrator agent creates tasks on the board (POST /api/board with title + status "todo")
2. Each coding agent in its CLAUDE.md is told: "Before starting any new work, POST /api/board/TASK-ID/claim. If you get 200, proceed. If you get 409, move to the next unclaimed task."
3. The server processes claims with SQLite's write serialization — exactly one agent wins each task, guaranteed, with no Redis, no distributed lock manager, no external dependency.
4. After a task is done and verified (CI green, deployed, smoke-tested), the agent marks it verified with the evidence. The board distinguishes `done` (implemented) from `verified` (confirmed in prod) — no premature completion.

Git conflicts are handled at the session level: each agent registers with its own branch (amux register feature-agent-1 --dir ~/Dev/myproject --yolo), and the CLAUDE.md template instructs agents to commit before switching context. Combined with git worktree isolation for truly independent parallel work, merge conflicts in parallel-agent repos are rare.

Self-healing overnight operation

amux's watchdog monitors every session for four failure modes:

• Context exhaustion: Detects the "approaching context limit" signal and auto-fires /compact before Claude Code's own resume dialog blocks the session
• Thinking-block corruption: Detects malformed thinking blocks in session output, restarts the process, and replays the last steering message
• Tool-approval blocks: In YOLO mode, the watchdog detects the unique "Esc to cancel" approval marker and sends the approval keystroke automatically
• Silent hangs: Idle timeout detection kicks unresponsive sessions after a configurable window

In practice, a 10-agent overnight run that previously required a developer to babysit (or produced incomplete work) now completes with 2-8 automatic recoveries handled by the watchdog — visible in the session's recovery log the next morning.

Multi-runtime rate-limit distribution

Hitting Claude Code rate limits is the single biggest throughput bottleneck for teams running parallel agents on Max plans. amux solves this structurally:

• Register Claude Code sessions against your Anthropic Max plan or API key
• Register Codex CLI sessions against OpenAI API keys
• Register Antigravity CLI (Google) sessions against your Google AI Pro subscription
• All three appear in the same dashboard, pull from the same kanban board, and coordinate via the same REST API

Distributing across three provider buckets gives 2-3× effective throughput compared to running all agents against a single provider. Route by task type: Claude for architecture decisions and code review (highest capability), Codex for standard feature generation (good quality, high volume), Gemini Flash for documentation and test generation (lowest cost — typically 30× cheaper than Claude Sonnet per token).

Built-in tools beyond coding

amux includes a set of built-in tools that make AI agents useful beyond writing code — important for engineering organizations that want agents handling the full developer workflow:

• CRM: Contacts with company, role, email, phone, LinkedIn, and interaction logs — AI agents can log customer meetings, track partnership conversations, or manage hiring candidates
• Gmail API email: Send, reply, and read email through any connected Gmail/Workspace account ([email protected], [email protected]) — agents can send PR review notifications, status emails to stakeholders, or weekly summaries
• Browser automation: A shared Playwright browser instance with saved auth profiles — agents can run competitor monitoring, scrape job listings, check production health pages, or fill forms
• Scheduler: Named recurring jobs (cron or natural language) — daily standup summaries at 7am, weekly dependency audits, monthly changelog drafts
• Notes: Shared markdown wiki accessible via REST API across all sessions — agents write documentation, research, and runbooks that persist across session restarts

Audit trail and governance

amux is append-only by default: every session's full JSONL transcript is stored, every board task transition is timestamped, and every schedule create/delete includes session attribution (AMUX-1812). This gives engineering leadership:

• A complete record of what every agent did, in what order, with what input
• Board issues showing who created each task, which session claimed it, and when it was completed and verified
• Schedule audit trail showing who created or deleted recurring jobs (session name or IP, timestamp)
• CLAUDE.md version control: your org-wide agent instructions are checked in alongside code, reviewed in PRs, and enforced across every session via the rules system

For SOC 2 or ISO 27001 purposes, the JSONL transcripts and board history provide the evidence trail for "what did the AI system do and who authorized it."

Deployment options

What a CTO actually sees on day 1

Here's a realistic first-week picture for a 5-developer engineering team that deploys amux on a shared server:

1. Day 1: Install amux on a team server (pip install amux, configure TLS, add auth). Each developer registers their project: amux register myproject --dir ~/Dev/myproject --yolo. The CTO and team leads get the iOS app and see all sessions on the dashboard within an hour.
2. Day 2: Create the first shared task queue on the board. Five developers add their current sprinted tasks as board issues. The first two developers with Claude Code sessions running claim and complete tasks while the CTO monitors from their phone in a meeting.
3. Day 3: First overnight run. 8 sessions start at midnight. The watchdog fires twice for context compaction, once for a crash recovery. In the morning: 6 of 8 tasks completed, 2 in doing (picked up second tasks after finishing their first). The CTO's morning board review takes 10 minutes from the iOS app.
4. Week 1: The team writes shared CLAUDE.md rules for their coding style and branching conventions, checked into the repo. All sessions now follow them automatically. Cost tracking in the dashboard shows $180 in Claude API spend for the week — CTO sets a per-session daily budget threshold in server.env.

Comparison: ad-hoc agents vs amux team deployment

FAQ

pip install amux
# or: brew install mixpeek/amux/amux
amux serve  # → https://your-server:8822

Related: amux for Enterprise · amux for Engineering Managers · amux for AI Engineers · amux for Startup Founders · Team Workflow Setup Guide · AI Agent Quality Gates