Let Opus Write Less Code: The Key to Saving Money Is When It Steps In
Anthropic’s advisor tool and Grok Bot’s agent org chart are talking about the same thing: strong models only make judgments, while execution is handed off to cheaper roles. The hard part isn’t building the chart; it’s the boundaries, acceptance checks, and knowing when to call a halt.
Anthropic’s docs include an experimental feature called the advisor tool. When starting Claude Code, add a parameter:
claude --model sonnet --advisor opus
The main model, Sonnet, does the work. Opus doesn’t take part in every turn; it is only called in at key moments to read the full conversation and give advice. The official docs are clear: advisor is a server-side tool that requires the Anthropic API and is unavailable on Bedrock, AWS, Google Cloud, and Microsoft Foundry. Claude decides when to consult it, tending to do so before committing to a plan, when errors keep recurring, and before declaring a task complete. The advisor reads the full transcript each time, is billed at the strong-model rate, and is not cached. The savings come from fewer calls, not from a lower unit price.
xsmryze’s tweet expands this into a division of labor: Sonnet 5.5 runs the main session and writes code, Haiku 5.5 serves as a read-only subagent that searches the repo, finds call sites, and pulls API docs in parallel, and Opus 5.5 acts as the advisor. He also provides a prompt that lets Claude Code rebuild the configuration as a tree, checking the diff before making changes. The video is below.
There’s a caveat here. The official docs only guarantee the advisor’s model pairing and invocation mechanism. Haiku subagents are another layer of orchestration and need to be tested yourself. In the model table, Sonnet 5.5 accepts Opus 5 or newer as an advisor; whether Haiku 5.5 can serve as an advisor depends on the version and pairing, so don’t assume it.
One reply said that, broken down, this trick is just using cheap models to do simple things. VeeraMarni said in a reply that there’s a new trick every week, and broken down, they’re all about using cheap models to do simple things. Another person said the key is when to bring Opus in, and that step still requires someone to judge. Frank said in a reply that he already acts as an adversarial reviewer, and waiting for the advisor’s reply could cost him negative six hours. These doubts are not unreasonable. The advisor’s benefit depends on two things: how often the strong model is called, and whether its advice is accepted and verified. Call it too often, and both the token bill and waiting time go up. Without acceptance checks, the advice is just another chunk of chat history.
smerkzz’s other article discusses Grok Bot, with an approach closer to a company org chart. A Chief of Staff takes goals, breaks down tasks, and gives one update per day. A Researcher scans sources. A Writer only writes drafts. A Reviewer only lists blocking issues. Ops gets added later. Each bot has its own goals, safety limits, quality bar, and acceptance checks. They share a task board. First run it manually, then save it as a skill, and only then put it on a routine. The article also warns that all bots share one computer; files, sessions, and logins are public property, so you need folders and handoff files to draw boundaries—but that’s organization, not a security boundary.
A comment said every box consumes the same plan, and other people can change the limits. Much of a chief of staff’s work is butler work and doesn’t need the strongest model. Fernando joked in a reply that when two experts’ opinions conflict, the chief of staff might just listen to the last answer. An org chart can reduce how often you act as the router, but it won’t automatically give you the right answer.
Looking at the two lines together, the common ground is role boundaries and acceptance checks; the number of models doesn’t matter. Roles should be narrow, handoffs should carry evidence, strong models should only make judgments, and acceptance criteria should be written first. The SimpleDot Agent Studio comment gives an engineering version: Kanban-driven, worker isolation, middleware for deterministic validation, dispatch, and collection, and on failure only BLOCK rather than automatically retry. Link here: SimpleDot Agent Studio. It solves observability and boundary problems for long-running multi-agent tasks; the model itself doesn’t get smarter because of it.
If you just want to try it, start with the minimal setup: a Sonnet main session plus an Opus advisor, run a long task, and record how many times the advisor was called, what it said each time, and whether it changed the result. Then add Haiku subagents. The same goes for Grok Bot: first build a Chief of Staff, give it a real task, and once that works, add Researcher, Writer, and Reviewer. Don’t build twenty bots on day one. The org chart is the result, not the starting point.
The models are someone else’s, the limits are someone else’s; only the acceptance criteria are yours. What saves money isn’t having many models—it’s calling strong models less often and calling them at the right times.
Reference links:
发布时间: 2026-11-07 20:31