Agent edits the workflow at build time, plain script at run time: token-saving browser automation
ZennoLab_Guru · reddit · 2026-09-29
The ZennoPoster team (a Windows browser automation tool) shared their architecture for using LLMs in exactly three places:
- Build time: an MCP client like Claude Code, Cursor or Codex edits the visual action graph — inspecting, adding and rewiring steps — after which the project runs as normal deterministic multi-threaded automation with no LLM involved.
- Repair time: when a task breaks, the agent inspects the graph, patches the failed branch, saves and restarts.
- Isolated run time: genuinely fuzzy steps embed an agent step with a strict iteration limit that must return RESULT: SUCCESS/FAILURE to pick the next deterministic branch.
Lessons: sending the full graph blows the context, so they expose structure at three levels (group overview / single group / full graph); agents once closed projects with unsaved changes, so closing now returns 409 unless discardUnsavedChanges: true; agents can check their permissions before starting instead of discovering denials mid-task.
More from coding & agent
- PHP WASM kernel hits major milestone with PR #936 merge, author vows to keep backing open source — MickeySteamboat · 2026-09-29
- Google to replace Gemini Gems with Skills starting November 17 — mark_k · 2026-09-29
- Carla v0.1.0: a local llama.cpp loom TUI for growing AI characters — max_paperclips · 2026-09-29
- Running Jev at high frame rate with full-state snap inferences makes it a true System 1 — mathemagic1an · 2026-09-29
- Meeting agent memory design: retain facts instead of patching tags — Extension-Answer4821 · 2026-09-29
- MegaMemory: a persistent knowledge graph MCP server that lets coding agents remember across sessions — tom_doerr · 2026-09-29