Building 24/7 agents: initial prompts add zero value, chat-based tuning wins
_rchaves_ · x · 2026-11-20
Quoted post: Kortix launched a tool where one prompt adds a "software factory" to your repo — self-starting agents running in loops that watch Slack, logs, tests, support inboxes and feature requests, turning findings into PRs; 40% of PRs merged in their 20k-star repo this month were agent-written.
Rafael Chaves pushes back: asking Claude to create a "24/7 support inbox agent" or "Sentry errors agent" is the easy part — the initial prompt adds zero value. The hard part is defining what's specific to your team and business.
His hands-on lessons:
- YAML-file configuration doesn't work; they tried it — one person sets it up, nobody maintains it, and the agent accumulates dust while eating tokens
- It has to be done via chat: multiple people, including non-technical folks, continuously nudging the agent as tiny issues come up — this has worked wonders for them
- Business-critical agents shouldn't use this vibed approach at all; they need evals and proper engineering, not a "software factory" tool
More from coding & agent
- Apple Pages, Numbers and Keynote add MCP support with a local server on Mac — jedisct1 · 2026-11-20
- Overmind: an open-source AI lab that turns agent traces into your own fine-tuned models — rohanpaul_ai · 2026-11-20
- Stacking Antigravity CLI under Claude Code: Opus 5.5 drives Gemini Neon — boringmarketer · 2026-11-20
- Kelpie: open-source browser agents can drive on real phones and desktops via MCP — Turbulent_Rooster_73 · 2026-11-20
- GlassFolio: self-hosted wealth dashboard with client-side AES-256 encryption and local AI — Dam_damm · 2026-11-20
- Bloomberg: Apple building its own autonomous AI agent, launch possibly late next year — kimmonismus · 2026-11-20