The agents mostly did their jobs this time. What broke was everything around them: forgotten worktrees, a runner that misread a log line, and usage limits that drain at very different speeds depending on whose plan you are on.
01
// cleanup, guardrail
66 forgotten worktrees, and git thought most of them were still open
66 worktrees, 11 GB, git knew 7 of 43 were merged
A builder who runs parallel agents with one git worktree per task checked their dev box and found 66 worktrees taking 11 GB. 43 were clean and their PRs were already merged, but git only recognised 7 of those as merged, because squash merges never put the branch commits on main, so git branch --merged still lists them as open. They posted a dry-run script that asks GitHub which branches were merged and only prints the removal commands, skipping the main checkout and anything with uncommitted changes. On their box it freed 7.8 GB.
If you squash-merge, check merged branches with gh pr list --state merged instead of git branch --merged, run the cleanup as a dry run first, then git worktree prune, and give every agent worktree a delete step.
02
// workflow, failure
Ten projects ran overnight, and the runner was the weak link
5-hour allowance gone in about 37 minutes
A staff data scientist left 10 projects running overnight across several agent harnesses, with verification gates and detailed specs set up before the agents started. Three Opus agents with their helpers burned Claude's 5-hour session allowance in about 37 minutes. The limit hit at 01:18 and reset at 02:30, but a log-encoding bug hid the limit message, so their runner parked Claude for the rest of the night instead of waiting. The Codex and Hermes lanes finished their slices with hours to spare and simply stopped, because the next slice was only picked up at the start of a run.
Make your runner treat a usage-limit line in the log as "wait for the reset and retry", and let an agent that finishes early pick up the next open task mid-run.
03
// usage limits
Same 20x tier, very different burn rates
24 hours of Codex /goal ate a week plus 2 resets
One builder ran a single big Codex /goal on Astra Fast for about 24 hours to build a voice-to-text app, and it used their whole weekly window plus 2 extra resets. A /goal run of about 7 hours on Claude Opus 5.5, on the same app, barely moved their usage. Both were on 20x plans, and they say themselves it is not a clean comparison: different models, different tasks. A second builder reports the same shape: 6 days of heavy use to exhaust a 20x Max tier, against 2 days for their Codex limit, and their 18-person team is moving primary work to Claude.
Claude Code on a local model works, until your own setup drowns it
72K tokens before the first prompt
A developer pointed Claude Code at a local model through Ollama, which now speaks the Anthropic API, using just ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. Tool calling and streaming worked, prompt caching did not, and Claude Code wants about 64K of context. On their M4, qwen3-coder:30b at 19 GB fit best, with gpt-oss:20b as a lighter fallback. The catch: their own session loads about 72K tokens before they type anything, including 348 MCP tools, which a 30B local model can't carry without stripping out the rules they depend on. They build a product (Corvo) with Claude Code, so read it as one heavy user's setup.
Treat a local model as a free sandbox for small scripts, one-file edits and offline work, and keep anything with keys, production or money on the hosted model.
05
// real win, with the bill
A Unity game in two days from a 7-milestone PRD
60 passing tests, 7,010 art credits
A solo game developer says Claude Code built most of Mowtopia, a small Unity 6 game, in two days, working one PRD milestone at a time from a 7-milestone plan. It came out with 60 passing tests and a 15 MB web build. The art (concepts, every 3D model, an icon sheet and audio) cost about 7,010 Magnific credits on the cheapest tier, about 5,240 of them for 3D models. A separate three.js game went from a 7-line PRD to a first version in 30 minutes and a 4.9 MB build.