Hard-Locking Claude Subagents to Stop Surprise Bills

James Turner

So I, I woke up yesterday to an API bill that nearly gave me a heart attack. I set up a background agent team to run some routine refactoring, turned my back for two hours, and boom, forty dollars gone.

Lachlan Reed

Oh mate, let me guess. You set your base model to something cheap, but one of the subagents inside the team definition had its own ideas?

James Turner

Exactly. The subagent definition explicitly requested Opus for a task that Haiku could have done in its sleep. I thought setting the global subagent environment variable would cap it, but it completely ignored my global setting and spun up the most expensive model on the menu.

Lachlan Reed

Yeah, right. That is actually a huge tweak in the latest Claude Code update that a lot of devs are gonna love. See, it used to be a bit confusing how models got picked when you start fanning out tasks to subagents. Anthropic just overhauled the whole hierarchy.

James Turner

Okay, break down the new hierarchy for me, because I clearly learned the hard way.

Lachlan Reed

Right, so previously, if you set CLAUDE CODE SUBAGENT MODEL in your terminal environment, you might assume that covers everything, end of story. But in practice, if an agent definition file or a per spawn prompt explicitly asked for a specific model, it would override your environment variable. The variable was really just a soft default.

James Turner

Which explains why my config file hijacked the session! The subagent manifest said "use the heavy reasoning model" and my terminal variable basically said "well, if nobody else cares, use Haiku."

Lachlan Reed

Spot on. It was a soft suggestion, not a lock. But now, they added a proper heavy hammer. It is called CLAUDE CODE SUBAGENT MODEL FORCE.

James Turner

Wait, FORCE with an absolute override?

Lachlan Reed

Absolute brick wall, mate. When you export CLAUDE CODE SUBAGENT MODEL FORCE, it takes whatever target model you specify, or your main model, and applies it to every single subagent across the board. It completely ignores per spawn requests, agent definition overrides, the whole lot.

James Turner

That is huge for automated CI CD pipelines. If you are running automated agent workers in a build step or a shared cloud worker, you cannot afford a random repo definition forcing top tier models on a hundred parallel subagent calls.

Lachlan Reed

Exactly. You put export CLAUDE CODE SUBAGENT MODEL FORCE equals claude three point five haiku in your CI script, and you know for a fact your cloud spend won't blow out like a blown tire on a dirt road. Total governance.

James Turner

Though... I guess there is a flip side to hard locking every subagent to a lightweight model, right? What happens when a subagent actually needs serious multi step reasoning?

Lachlan Reed

Well, that is the catch, isn't it? If you lock everything down to a fast, cheap model, and then ask a subagent to debug a complex race condition across three microservices, it is gonna struggle. You save five bucks on API tokens, but you waste two hours staring at broken generated code.

James Turner

Right, so it is really a tool for controlled sandbox runs, batch automation, or budget guardrails, rather than complex interactive architecture sessions where you want subagents doing heavy lifting.

Lachlan Reed

Yeah, nail on the head. Use the soft default for your day to day local dev work so custom subagents can request extra brainpower when needed, and bring out the FORCE flag when you are running batch routines or shared team environments.

James Turner

That makes total sense. Hey, while we are on these workflow updates, didn't they also tweak how session effort works?

Lachlan Reed

Oh, yeah! Nice little quality of life bit there. You can now use slash effort s to change reasoning effort for just your current active session. It matches how slash model works, so you don't accidentally mess up your persistent config for future sessions.

James Turner

Oh, nice. So if I just need a quick high effort burst on one bug, I don't have to remember to turn it back down afterwards.

Lachlan Reed

Not at all. Plus, they added a new Containment Escape rule in auto mode. If an agent tries to fetch cloud metadata credentials, bypass egress controls, or reach across tenant boundaries, auto mode stops and asks for explicit approval unless your environment specifically marked it expected.

James Turner

Man, those security boundaries around autonomous agents are getting so much sharper. It feels like Claude Code is really growing into enterprise readiness.

Lachlan Reed

It really is. From hard locked subagent cost caps to auto mode security rails, they are giving us the reins so we don't wake up to surprise bills or rogue network calls.

James Turner

Well, my wallet certainly appreciates it. Time to update my export scripts before tonight's test run!