GitHub user odabir33 added a comment to the discussion: Task ideas for the dkNet-AI · Apache Texera Agent Hackathon
## Idea 1: Plan-then-Build: draft before building **Frustration:** The agent jumps straight to creating operators. When it guesses wrong, you watch 6+ wrong operators appear, pay tokens for each, and regenerate the workflow or clean up manually. **Fix:** Insert a text-only plan phase before any mutation. `Request → Clarify questions → Markdown plan → Approve → Build` The agent emits a short plan (operators, params, assumptions, open questions) and waits. User refines in plain English (*"skip step 3"*). Nothing touches the canvas until **Approve**. **Why this matters?** - Wrong plans cost a 10-second read, not a 5-minute cleanup - Skips workflow mutation, recompilation, and per-step context rebuilds — one cheap call instead of many expensive ones - Clarifying questions help novices articulate vague intents --- ## Idea 2: Sample-then-Scale: try it on 100 rows first **Frustration:** *"I waited 8 minutes for it to crash at op12. Now I wait 8 more."* Most errors (schema, join keys, filters) show up instantly on a sample. **Fix:** Default the first run to **100 rows** per source. User reviews, tweaks, then clicks **Run on full data**. `Build → Sample run (~3s) → Review → Full run` **Why it matters?** - Catches ~90% of bugs in seconds, not minutes - Just inject `TableLimit(100)` on sources at submit; strip for full run - Stacks with Idea 1: plan → build → sample → scale --- Idea 1 catches *intent* errors before any work happens. Idea 2 catches *data* errors before any waiting happens. GitHub link: https://github.com/apache/texera/discussions/5059#discussioncomment-16925445 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
