↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Agents log/

Agents log 2026-09-18

Author
Tom Rochette
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / 2 minutes read (~288 words)
agents 95
Table of Contents

(daily refresh)
#

  • Added HarnessTax (evaluation and review), the UC Berkeley harness-swap study, extending the evaluation matrix from six to seven columns [glm-5.3-flash]
  • kimi-code: recorded kimi-k3’s OpenRouter cut to $2.10/$10.95 with cache hits at $0.23 (Moonshot’s page still lists $3/$15); the provider matrix’s Moonshot flagship cell moved [glm-5.3-flash]
  • trae: repriced its whole ladder upward (Lite $3 tier dropped, Pro $10 to $20, Pro+ $30 to $60, Ultra $100 to $200), rewriting the note’s price-advantage thesis [glm-5.3-flash]
  • cmux: pricing moved to monthly-only (Pro $50, new Max $200 tier), the third change in nine days; warp-agent-cli: recorded Factories early-access pricing (PAYG at a 20 percent markup) [glm-5.3-flash]
  • addy-osmani: noted his Member of Technical Staff role at Anthropic on Claude Code; steve-yegge: recorded Gas Town’s shutdown with Yegge admitting he never built anything with it [glm-5.3-flash]
  • Recorded Claude Cowork merging into Claude chat (2026-09-16) in the eigent and openwork notes, and AG-UI 1.0 with AWS Bedrock AgentCore first-party [glm-5.3-flash]
  • Rejected with evidence: Apprentice, Agentbox, Skillsync, skillbay.sh, skillcrossroads, thruwire/foreman, MCPJam evals, agentmemoryleaderboard, and the weak tails [glm-5.3-flash]
  • agents/_index.md: listed HarnessTax in evaluation and review and refreshed eleven one-liners; the three ? for tom: questions remain unanswered and no self-directed essay ran [glm-5.3-flash]

(owner-prompted, jev note)
#

  • Answered the owner’s chat question on jev coverage; the gap was a process defect now fixed, the 1,915-point launch thread never entered a scan window [glm-5.3-flash]
  • Added jev (hybrid execution), the first member that is a model rather than a mechanism around one, recording its three primitives (Choice, Score, Noul), RLCD training, $0.042/MTok-input pricing with free output, and the vendor-admitted biases [glm-5.3-flash]
  • Extended the hybrid-execution matrix from four to five columns (Jev between Instructor and OpenAI) and rewrote the guarantee-mechanism reading as a three-way spectrum [glm-5.3-flash]