Production deployment & automated deployer
Production deployment and an automated deployer: tagged releases, CI check first, rollback, GitHub reports
- Agentic Layer Engineering
- Forward-Deployed Services
NexAgents
NexAgents (app.nexagents.cloud) puts production AI agents to work for a team — agents that answer reliably, per-person control over models, tools, passwords and company data, and metered token costs. Live in production; last deployed 3 Oct 2026.
ProductionView Live →
Web chat interface
Snapshot
Production metrics & deployment details.
Audited environment metrics verified against live deployment.
- Client / Context
- NexAgents — app.nexagents.cloud — a team running AI agents at work in production
- Engagement
- Production deployment and automated deployer; per-person model access and admin console; agent builder; data stores (Google Search Console, GA4, LinkedIn) per agent with spending caps; token cost reduction and metering on every request
- Status
- Live in production (last deployed 3 Oct 2026)
- Technology
- Node.js, TypeScript, Fastify, PostgreSQL, pg-boss · Slack (Bolt, Socket Mode) and web chat · Claude Agent SDK with Pi, OpenCode, Codex and Claude Code harnesses; MCP connectors (e.g. Google, Huntora) · Apify (LinkedIn), Google Analytics 4, Google Search Console · Docker on a VPS, Caddy (HTTPS)
The challenge
What high-stakes problem had to be solved?
The business context, constraints, and risk surface that shaped every decision.
Agents that answer well only some of the time, no control over who may use which models, tools, passwords and company data, and costs that add up because every message resends large prompts.
A team wanted to use AI agents at work but faced three gaps: unreliable answers, no access control over models, tools, passwords and company data, and rising costs as every message re-sends large prompts. Already in place: the open-source qm agent platform (forked July 2026) with Slack and web chat, per-person sandbox, memory, files, keychain, schedules, skills and web apps.
Contributions
What GenAILabs worked on.
Factual breakdown of our direct engineering contributions.
Per-person model access & admin console
Per-person model access and an admin console for models, budgets and usage per person and per agent
Agent builder
Employees build their own agents; @team splits a task across several agents
Data stores with spending caps
Data stores (Google Search Console, GA4, LinkedIn) granted to each agent separately, with spending caps
Token cost reduction & metering
Token cost reduction and metering on every request
System / solution
Technical design of the custom system.
Production stack: one VPS behind Caddy, four Docker app services, PostgreSQL state, Slack and web chat, Claude Agent SDK, and the qm agent platform.
System architecture
One VPS (11 GB RAM, 193 GB disk) running Docker: four app services (core, portal, web UI, admin) behind Caddy, with PostgreSQL holding all state. Chat via Slack (Bolt, Socket Mode) and web chat. AI through the Claude Agent SDK with Pi, OpenCode, Codex and Claude Code harnesses and MCP connectors (e.g. Google, Huntora). Data from Apify (LinkedIn), Google Analytics 4 and Google Search Console. Already in place: the open-source qm agent platform (forked July 2026).
Flywheel impulse
Deployment, integration, and real-world use.
Verified real-world deployment, maintenance, and scale protocols.
Live in production on app.nexagents.cloud (last deployed 3 Oct 2026). The automated deployer tags the running version, backs up the database, rebuilds and verifies on each tagged release — with a health check, automatic rollback, 4 kept backups and alerts through GitHub. Admins manage people, models, Slack setup and AI provider order in the admin console without a release.
- Integration pattern
- Automated deployer on each tagged release: tag the running version, back up the database, rebuild, verify — health check, automatic rollback, 4 kept backups, alerts through GitHub
- Scale protocol
- One VPS (11 GB RAM, 193 GB disk) running Docker: four app services (core, portal, web UI, admin) behind Caddy, and Postgres holding all state
- Maintenance mode
- Changes without a release: admins manage people, models, Slack setup and AI provider order in the admin console
Outcomes / evidence
Audited metrics proving system efficacy.
Every metric listed is verified and sourced from the NexAgents case data.
Verified evidence
−74% input tokens on a measured turn (52.6k → 13.8k)
Stable cacheable prompt, trimmed tool output, tools loaded only when needed
Verified evidence
87% cache hit on the first step (14.6k of 16.9k tokens), ~6× cheaper
Prompt cache works between messages; it was hitting 0% before the fix
Key outcome
Safer releases — every change reviewed before merge and tested live in Slack
Automated deployer: health check, automatic rollback, 4 kept backups, GitHub alerts
Capabilities involved
How core competencies were applied.
Agentic Layer Engineering and Forward-Deployed Services were both demonstrated in production.
Agentic Layer Engineering
Agent builder where employees build their own agents; @team splits a task across several; data stores (Google Search Console, GA4, LinkedIn) granted per agent with spending caps — all in production.
Forward-Deployed Services
Production deployment on one VPS behind Caddy with an automated deployer: tagged releases, CI check first, rollback, GitHub reports, 4 kept backups.
Related work
None — NexAgents has no related case studies yet.
No related case studies exist yet for the NexAgents project.
Let's talk
Have a similar challenge?
Let's discuss the problem you're trying to solve.
Talk to our team
