Building a local Gemma planner-executor-critic agent pipeline for web research
HlddenDreck · reddit · 2026-09-03
A developer shares a fully local agentic web-research workflow built on opencode with 4 agents: Gemma-4-31B-IT-QAT (120k context) handles planning, reviewing and orchestration, while Gemma-4-12B-IT-QAT (256k context) executes. The planner runs initial web searches to vet sources, splits the plan into batches, and the orchestrator invokes the executor per batch to keep context small and avoid compaction. A critic checks sources and loops back to the planner for fixes (max 3 rounds). It can beat running everything on the 31B model (8-10 min per task), but is unstable — results aren't reproducible and tasks sometimes balloon to an hour. The author is asking for others' experience with similar pipelines.
More from coding & agent
- Non-coder runs hundreds of thousands of AI-written lines on a phone via zero-trust Termux workflow — amadale · 2026-09-03
- Multi-agent debugging pain: per-agent logs can't show which output changed the next agent's decision — mageblex · 2026-09-03
- PR adds ready-made JS and Python computer-use environments to OpenAI's CUA sample app — ChrisGPT · 2026-09-03
- 2-hour vibe-coded local agentic browser clones Opera Neon's $19/month features with MCP — comperr · 2026-09-03
- ComfyUI plugin caches MiniMax H3 conditioning to disk: 1.12s cache hits, 14GB RAM saved — Any_Fee5299 · 2026-09-03
- Open-source coding agent Pi hits 100,000 GitHub stars, v2 teased — gklambauer · 2026-09-03