Like others the agents are very slow to respond. I'm on a hostinger VPS.
Versions tested:
- 2026.4.27
- 2026.5.2
Measured results:
- 2026.4.27: avg wall 25.9s, avg agent 14.3s. Logs showed event-loop/cpu warnings, max event-loop delay about 8.5-9.0s.
- 2026.5.2 after config repair: avg wall 16.2s, avg agent 10.1s, steady wall excluding first 14.7s. First turn about 30s.
- 2026.5.2 after slimming Odin AGENTS.md to 5,750 bytes: avg wall 14.4s, avg agent 8.4s, steady wall 13.5s. First turn improved to 22.4s.
- Odin tools.profile="messaging": did not help. Avg wall 17.7s, steady wall 14.9s, avg agent 9.3s. Reverted.
Confirmed:
- Upgrade helped: 25.9s avg wall to 16.2s.
- Slim AGENTS.md helped further: 16.2s to 14.4s.
- Plain chat still usually 8-15s.
- Tool/diagnostic turns can still be 20-40s.
- One 26s turn happened before any multi-agent burst. It was a tool-using diagnostic turn, including a failed
pythoncommand because onlypython3exists. Simple follow-ups in that window were about 8.4s. - USER.md existed but was symlinked outside workspace. OpenClaw marked it missing. Replacing symlinks with real small files fixed injection. This affects context, not latency.
Remaining bottleneck:
core-plugin-tool stagesrepeatedly take about 2.3-2.6s per turn.- Most is
openclaw-tools:pdf-tool, about 2.1-2.9s. - I do not use PDFs with these agents.
- Added
tools.deny=["pdf"], validated, restarted. pdf-toolstill prepared and still cost about 2.3s.tools.profile="messaging"still left 80 tools and 71k schema chars.
Questions:
- How should I reduce MCP/plugin tool schema load per agent?
- Is 8-15s steady Discord response expected for openai-codex/gpt-5.4 on 2026.5.2?
Any way to get it lots faster?