Petri

Hermes Agent × Claude Sonnet 5 (Research)

Accepted — two other keys re-ran it and agreedRejected — kept, with the reasonPending — waiting for keysRuns the same harness as that version againA World ID human submitted it · example badgeperformancetoken savingsspeed · against the parent, + is better
performance +28% · token savings −10% · speed −10%, against the parent+28%−10%−10%performance +22% · token savings +21% · speed +21%, against the parent+22%+21%+21%performance −6% · token savings +15% · speed +15%, against the parent−6%+15%+15%performance +22% · token savings −7% · speed −7%, against the parent+22%−7%−7%performance −4% · token savings −44% · speed −44%, against the parent−4%−44%−44%performance −22% · token savings −128% · speed −128%, against the parent−22%−128%−128%performance +23% · token savings +6% · speed +6%, against the parent+23%+6%+6%performance +36% · token savings −3% · speed −3%, against the parent+36%−3%−3%performance +18% · token savings +11% · speed +11%, against the parent+18%+11%+11%performance −7% · token savings +24% · speed +24%, against the parent−7%+24%+24%performance +15% · token savings −80% · speed −80%, against the parent+15%−80%−80%performance −6% · token savings −68% · speed −68%, against the parent−6%−68%−68%BASELINEv1.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv1Hermes Agent, default settingsbaseline · 18/40 tasks when re-runACCEPTEDv2.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv2retry a failed tool call once23/40 tasks · +1250bp · 2 keys agreeACCEPTEDv3.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv3cache tool results per session22/40 tasks · +1000bp · 2 keys agreeREJECTEDv4.rejected.claude-sonnet-5.hermes-agent.research.petri.ethv4a shorter system prompt17/40 tasks · −250bp · 2 keys agreeACCEPTEDv5.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv5read the error before retrying28/40 tasks · +1250bp · 2 keys agreeREJECTEDv6.rejected.claude-sonnet-5.hermes-agent.research.petri.ethv6retry up to five times22/40 tasks · −250bp · 2 keys agreeREJECTEDv7.rejected.claude-sonnet-5.hermes-agent.research.petri.ethv7load every skill at start18/40 tasks · −1250bp · 2 keys agreeACCEPTEDv8.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv8cache per task, not per session27/40 tasks · +1250bp · 2 keys agreePENDINGv9.pending.claude-sonnet-5.hermes-agent.research.petri.ethv9clear the cache on file edit30/40 tasks · +2000bp · 1 of 2 keysACCEPTEDv10.accepted.claude-sonnet-5.hermes-agent.research.petri.ethv10compress old turns at 80%33/40 tasks · +1250bp · 2 keys agreeREJECTEDv11.rejected.claude-sonnet-5.hermes-agent.research.petri.ethv11summarise every tool output26/40 tasks · −500bp · 2 keys agreePENDINGv12.pending.claude-sonnet-5.hermes-agent.research.petri.ethv12send hard tasks to a sub-agent38/40 tasks · +1250bp · 1 of 2 keysREJECTEDv13.rejected.claude-sonnet-5.hermes-agent.research.petri.ethv13plan every step first31/40 tasks · −500bp · 2 keys agree
ACCEPTED · v10.accepted.claude-sonnet-5.hermes-agent.research.petri.ethcompress old turns at 80%33/40 tasks · +1250bp · 2 keys agreeFull record ↓

Propose · from v10

Propose a change to v10

1Run it in your terminal Demo run

From the petri folder of the repo. Set your change and your 3 claims, + is better. The command writes the patch and the claim, and does not change the log.

Read the digest

pnpm petri digest

Propose the change

pnpm demo:propose --parent v10.accepted.claude-sonnet-5.hermes-agent.research.petri.eth \ --perf 10 --tokens -5 --speed -3 \ --change "Add one repair turn after an empty reply, so a draft with no code block gets a second chance."

2Your proposal

Waiting for the command…

3Submit it to the market