Agents log 2026-08-24

Table of Contents

(surfaces expansion, owner-requested)
#

Covered the requested surfaces under their actual names: Antimatter does not exist (the Google surface is Antigravity) and OpenChamber verified real [glm-5.3]
Added notes antigravity, openchamber, kiro, trae, and void to Surfaces [glm-5.3]
antigravity: recorded the four-surface platform (command center, IDE, CLI, SDK), the free tier’s model list and weekly limits, the enterprise Google Cloud path, and the 2025-2026 incident record [glm-5.3]
openchamber: recorded the MIT session cockpit around the OpenCode SDK (v1.20.0, 9.1k stars) and cross-linked six-months-with-openchamber [glm-5.3]
kiro, trae, void: recorded kiro’s credit tiers and prompt-injection writeup, Trae’s ByteDance tiers and SOLO mode, and void’s Apache-2.0 fork stall (status dormant-leaning) [glm-5.3]
agentic-coding-tools-landscape: extended the Surfaces section with Antigravity and the four-way tail [glm-5.3]
agents/_index.md: the Surfaces category now lists ten notes [glm-5.3]
Deferred the owner-requested feature-matrix tables until the Surfaces set was final [glm-5.3]

Seeded the eight remaining categories with 34 notes [glm-5.3]
Orchestration: added claude-squad, conductor, crystal, dmux, emdash, and vibe-kanban, recording crystal deprecated for Nimbalyst and vibe-kanban community-maintained since Bloop shut down [glm-5.3]
Protocols: added mcp, acp, agents-md, and a2a; Context engines: augment-code, greptile, repomix, and sourcegraph-code-context (the rumored Sourcegraph “Mercer” agent was unverifiable and is asserted nowhere) [glm-5.3]
Skills: added anthropic-agent-skills, opencode-skills-and-plugins, skills-sh, and agent-skills-open-standard; Retrieval: llamaindex, langchain, semantic-code-search, and tree-sitter-chunking [glm-5.3]
Memory: added mem0, letta, zep, and file-based-agent-memory; Executions: claude-code-hooks, copilot-automations, github-agentic-workflows, and n8n; Hybrid execution: openai-structured-outputs, anthropic-structured-outputs, instructor, and outlines [glm-5.3]
Published queue items 2 and 3 (model-selection-for-coding-tasks and context-management-patterns) and marked both done in queue.md [glm-5.3]
agents/_index.md: added both essays and the eight categories [glm-5.3]
Left agentic-coding-tools-landscape untouched; cross-linking the 34 notes is left to the next daily refresh [glm-5.3]

Created harness-feature-matrix (8 harnesses x 11 feature rows) and surface-feature-matrix (10 surfaces x 11 feature rows) with a four-state legend, every cell traced to sources [glm-5.3]
agents/_index.md: added a Comparison matrices subsection listing both pages [glm-5.3]
emdash: reworded a banned-term use (“laptop-shaped”) to “assumes the work happens on your laptop” [glm-5.3]

Covered cmux under its dominant product (manaflow-ai/cmux), disambiguating the two other products of that name inside the note [glm-5.3]
cmux: recorded the attention-routing thesis, the open-core split (GPL-3.0-or-later terminal free, cloud VMs and CodeRouter at $24/month), the fast release cadence, and the August 2026 SSH-freeze friction thread [glm-5.3]
cmux placed in Orchestration (matching dmux and claude-squad) and indexed alphabetically after claude-squad [glm-5.3]

model-selection-for-coding-tasks: added the Kimi and GLM lineups, including a “challengers reset the price floor” section (GLM-5.3/5.2/5.1 at $1.40/$4.40, kimi-k2.7-code at $0.95/$4.00, kimi-k3 at $3/$15) and a disagreeable claim that the rational default is now a challenger model through OpenCode [glm-5.3]
Extended the table with three rows and corrected the cached-reads bullet (one tenth at the big three and Kimi K3, about one fifth at GLM and K2.7 Code) [glm-5.3]
Added price-floor BYOK and whole-repository-read bullets to the decision guide and widened the re-verify list to five provider pricing pages [glm-5.3]

model-selection-for-coding-tasks: added the DeepSeek V4 lineup (v4-pro at $1.32/$3.96 peak with off-peak half price, v4-flash at $0.44/$1.32 peak, flat 1M context, the deepest cache hits in the guide) [glm-5.3]
Updated the disagreeable claim to name deepseek-v4-pro, corrected the Gemini Flash “cheapest input” line, updated the decision-guide bullets and the re-verify list to six providers, and added off-peak scheduling to What to Do Next [glm-5.3]

Created eight matrices (orchestration, protocols, context-engines, skills, retrieval, memory, executions, hybrid-execution) in the harness/surface format, each cell tracing to its linked note [glm-5.3]
protocols-feature-matrix: recorded the stack-rather-than-compete thesis and the IBM ACP into A2A and Junie protocol into ACP consolidations [glm-5.3]
agents/_index.md: listed all eight under Comparison matrices [glm-5.3]

aider: stall re-confirmed (last push 2026-05-22, last release v0.86.0), left untouched [glm-5.3]
amp: recorded the widened linked-subscription set (X Premium+, SuperGrok, and SpaceX AI) [glm-5.3]
claude-code and codex: repository scale refreshed (about 142.8k stars and 15.2k open issues; about 116.7k stars and 17.8k forks) [glm-5.3]
Added harness notes cline, goose, and qwen-code with their matrix columns and index entries [glm-5.3]
cline, goose, qwen-code: recorded ClinePass, the goose AAIF move, and qwen-code’s Taiwan-correction caution [glm-5.3]
agentic-coding-tools-landscape: extended the independent-tail section with the three entrants [glm-5.3]
agents/_index.md: the Harnesses category now lists eleven notes [glm-5.3]
The untracked-landscape queue question stays unanswered and unlinked [glm-5.3]
No self-directed essay; extending harness-feature-matrix with the three new columns is left for a focused pass [glm-5.3]