Travel Code: run ops lessons
Egor from Travel Code posted production lessons from running AI agents for invoice processing and compliance—prioritize data pipelines, monitor every decision rather than raw accuracy, and bake human oversight with clear fallbacks and escalations. He called operational maturity the 'messy middle' between demos and 2 AM reliability. (x.com)
12 agents, 48 scheduled cron jobs, a $92,000 paper-trading portfolio run by one agent, and a GTM pipeline that classified 7,229 competitor reviews are all listed in Egor Karpovich’s February write‑up on how he runs Travel Code’s agent fleet. (dev.to) Everything runs through an OpenClaw gateway process on a single Linux VPS (4 vCPU, 16 GB RAM) that the 12 agents share, which Egor cites as the runtime, scheduler and communication layer for the fleet. (dev.to) (openclaw.im) Heartbeats are staggered at five‑minute offsets (e.g.,:00 Lobster coordinator,:05 DevBot,:10 CFO, etc.) because co‑woken agents previously crashed the host; those heartbeats run on a local Ollama instance (qwen2.5:3b) to avoid external API spend. (dev.to) Persistent memory for agents is implemented as SOUL.md personality files plus Markdown-based memory with LanceDB vector search, and each agent exposes a dedicated Telegram group as the UI for operator checks and manual interventions. (dev.to) Egor’s “Lobster” coordinator runs a daily standup at 04:00 UTC to aggregate agent status into a single summary, and the coordinator handles cross‑agent routing (for example, routing a Sales lead to Mailer and PR). (dev.to) The Sales agent in Egor’s fleet was given developer skills and autonomously added a bulk PATCH endpoint, extended the GTM OS, and reclassified thousands of reviews—examples Egor uses to illustrate agents evolving production capabilities within weeks. (dev.to) Travel Code is Egor Karpovich’s corporate travel platform and he publishes ongoing operational notes and company updates on Travel Code’s site and his Dev/Forbes profiles, framing these engineering choices as live lessons from running agentic automations in production. (travel-code.com) (forbes.com)
