Fable's judgement
Agent savings can come from fewer bad loops, not only cheaper tokens.
One of the most interesting tips I got from the Fireside Chat I hosted with Cat Wu and Thariq Shihipar from the Claude Code team at AIE on Wednesday was …
01
AI token economics, written like a field brief.
A bright, readable guide to what AI really costs: model prices, coding-agent workflows, benchmark signals, and practical ways to spend fewer tokens.
明快、可读、可追踪：帮个人和小团队看懂模型价格、Agent 工作流、评测信号和省 token 方法。
Price WatchInput, output, cache, batch, context, retry.
Agent CostClaude Code, Codex, Cursor, Aider, OpenCode.
Token SavingPrompt caching, routing, compression, context discipline.
Starting hooks
From top to bottom, each card gives title, date, summary, and the opening signal so readers can decide fast.
02
Release: llm-coding-agent 0.1a0
A minimal coding agent maps where token spend happens.
A coding agent built on LLM
03
Using DSPy to evaluate and improve Datasette Agent's SQL system prompts
Prompt optimization can be evaluated with harnesses instead of vibes.
Leveraging the DSPy framework, this project evaluates and refines the core production system prompts used by Datasette Agent’s read-only SQL question answerer. The methodology involves a harness where DSPy agents …
04
Vercel's Andrew Qu on why agents are a new kind of software
Agent-readable websites are becoming part of the product surface.
The Vercel Chief of Software explains how its agent framework, eve, was created — and why skills, sandboxes and agent-readable websites now matter.
05
How Cursor deploys AI inside the enterprise
Vibe coding becomes a team budget problem when workflows scale.
Cursor's Pauline Brunet explains how her team of Forward Deployed Engineers help organizations implement agents — essentially setting up software factories.
06
What’s new in Claude Sonnet 5
New model releases affect defaults, agent costs, and failure rates.
Claude Sonnet 5 came out this morning. I always head straight for the "what's new" developer docs because they tend to have more actionable information than the official announcement post. …
07
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
Agent benchmarks are moving toward real enterprise migration tasks.
A Blog post by IBM Research on Hugging Face
08
Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding
Open-weight coding models can change the API cost equation.
This is an interesting new open weights (MIT licensed) model, the first model release from DeepReinforce. [...] with variants including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. Built …
09
OpenAI reports median internal Codex output tokens grew dramatically
Output token growth is a major hidden cost in agent workflows.
It's happening.
10
GLM-5.2 is the step change for open agents
China/open models are part of global agent cost/performance comparisons.
A capability threshold I've been carefully monitoring.
11
Is it agentic enough? Benchmarking open models on your own tooling
Your own tooling may matter more than public leaderboard rank.
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
12
Qwen3 benchmark results
A durable bridge between China models and coding-agent evaluation.
Benchmark results for Qwen3 models using the Aider polyglot coding benchmark.
13
How Claude Code uses prompt caching
Prompt caching directly changes speed and token cost.
Claude Code manages prompt caching automatically. See why a model switch triggers a slow uncached turn, what /compact costs, why CLAUDE.md edits don't apply mid-session, and how to check your cache hit rate.
14
Ditching Claude for OpenCode and OpenRouter
A real switching case from default tools to open router/model workflows.
For the entirety of June I ditched Claude Code and have been using open weight models with Opencode and openrouter.ai. Here
15
Contextify - Searchable History for Claude Code and Codex
Agent history and reusable context can reduce repeated token spend.
Your Claude Code and Codex history auto-deletes. Contextify keeps it forever in a searchable database, syncs it across every machine, and runs on macOS and Linux.
Readable by people and agents
Static HTML first, with machine-readable endpoints for automation and search.
- /llms.txt
site purpose and reading policy - /feed.xml
RSS for briefs and hooks - /data/hooks.json
structured article cards - /sources/
source and citation policy