Open-Source sopc: Manage Prompts Across Many Voice Agents Like Code
BadAthMOFO · reddit · 2026-10-07
The author's team runs one voice agent per restaurant customer with mostly copy-pasted prompts—changing an allergy-handling rule meant hand-editing every prompt, and some had already drifted ("upsell at most once vs twice per call"). He open-sourced sopc (Apache-2.0) to manage this like code:
- Structure: shared instructions (brand voice, identity, policies, lockable), SOP procedures (goal, steps, never-dos, warning signs, tools), and one short per-agent file (platform ID, own facts, block references).
- CLI workflow: sopc plan shows which agents a change touches with exact prompt diffs before merge; everything is Markdown+YAML in git—review is a PR, rollback is a revert.
- Platform-agnostic: output is one prompt per agent with IDs for LiveKit, Vapi, ElevenLabs, and Retell; sopc verify catches dashboard edits drifting from git.
- Eval gating: CI runs evals (Braintrust, replay scripts, anything) only for agents a PR actually changes and blocks merges on regression.
- A single Claude Code command imports existing prompts and splits out shared blocks, verifying nothing is lost.
More from coding & agent
- Consultant ditches custom voice agent build, deploys Grok bot with Cursor teams — alexcovo_eth · 2026-10-07
- Benchmark: OpenAI Decisions API costs 2x more, 5-10% worse than Jev — xeophon · 2026-10-07
- Devs are cloning Photoshop in Rust and open-sourcing it — and AI-native tools need rethinking — ravisparikh · 2026-10-07
- Converting 10 coding harnesses including Claude Code into RL environments, cutting tool calls 31% — vllm_project · 2026-10-07
- AI-built websites are starting to look the same, developers blame shared design.md workflows — MickeySteamboat · 2026-10-07
- A mental model for AI assistants: dots own responsibilities, codex takes tasks — pvncher · 2026-10-07