Parallel coding agents tested: token bloat and coordination overhead
TheJoyfulTater · reddit · 2026-09-02
The author evaluated multiple parallel coding agent setups to improve efficiency, with mixed results:
- Claude Code + Superpowers: Sub-agents often failed to finish tasks, requiring main agent corrections, leading to context bloat and low efficiency.
- Claude Code Agent Teams: High token usage due to direct agent-to-agent chat. Switching to cheaper models (DeepSeek) saved money but not efficiency.
- OpenCode + omo-slim: Required detailed AGENTS.md for delegation; coordination still consumed significant context.
- OpenCode + omo + GPT 5.6: Full delegation (ULW) caused excessive delegation, rapid token drain, and context pollution with coordination messages.
- Codex + Matt's skills (Current): Uses a serial workflow (grill-with-docs -> to-spec -> to-ticket -> implementation). Good code quality, but parallel ticket processing still requires manual worktree/PR management.
The author seeks advice on optimizing workflows or model choices.
More from coding & agent
- After LLMs refused to crack DRM, an agent scripted screenshot+OCR to translate books — evisoft · 2026-09-02
- Fable 5.1 builds a macOS animation app in under 20 minutes — amos_gyamfi · 2026-09-02
- Testing AI Agents: Complement Not Replacement for Scripts — PartyVermicelli1870 · 2026-09-02
- Experimenting with Claude to Revive streamlit-drawable-canvas — andfanilo · 2026-09-02
- Would developers pay for an 'infrastructure layer' that equips AI agents with phone, email, and memory? — Athlore_AI · 2026-09-02
- Dev notes record outcomes, but the reasoning dies in the transcript — Sea-Perception1619 · 2026-09-02