One day with Opus 5.5 Ultracode: 28 subagents, 5.4M tokens, and endless review loops
mrbillyr2 · reddit · 2026-09-29
A developer vents about a full day building with Opus 5.5 Ultracode: every spec triggers arguments plus 2-3 subagents running reviews at each step, adding excessive guardrails and scope creep — "like hiring an army of graduate developers."
His session log shows:
- 28 subagents launched in one day, totaling 4h 17min of agent time and 5.4M tokens
- A single spec workflow (draft → verify-code-truth → verify-buildability → revise → backlog) took up to 81 minutes wall-clock, with one draft running 35 minutes and 482k tokens alone
- Even fact-checking a features guide spawned 4 parallel checker agents
He now feels like a technical presales/product manager just steering agents, suspects his many installed Matt Pocock skills, and says he'd rather go back to Opus 4.6 and GPT 5.4.
More from coding & agent
- The hidden primitive behind Claude Code, Codex and Gemini: verifiers agents can't see — bibryam · 2026-09-29
- Solo dev weighs $5k-$20k local AI upgrade: dual RTX 3090 rig vs clustered DGX Spark GB10 — ChopSticksPlease · 2026-09-29
- Reddit thread: how should agents log the decisions where they choose not to act? — daani_maas · 2026-09-29
- On open-ended bug-fixing tasks, Astra averages 222 turns vs Sonnet 5.5's 1330 — PawelHuryn · 2026-09-29
- Claude Code user wants a default-model setting for workflow subagents after runaway 200-agent session — Liu_eroteme · 2026-09-29
- When Shopify WebMCP checkout disagrees with the visible UI, who should the agent trust? — daani_maas · 2026-09-29