--- title: "Overpowered subagents" description: "antiburn flags sessions where a premium main model delegates work to subagents running on that same premium tier." url: "https://antiburn.ai/docs/findings/overpowered-subagents/" --- ## What it detects When a session's main-loop model spawns subagents, antiburn checks the tier of both: the dominant main-loop model (the model doing most of the main-thread turns) and every model the session delegated work to. A finding fires when the main model is on a reviewed premium tier - Claude's `opus`, `fable`, and `mythos` models, or OpenAI's top-tier `gpt-6-astra` and `gpt-5.5`/`gpt-5.6` variants, or a Gemini `pro` model - and at least one delegated model is also premium. Premium status is judged per model family, using a maintainer-reviewed list, not a heuristic - a family antiburn hasn't reviewed (or an unrecognized model) can never prove premium or non-premium, so it never produces a false finding. ## Why it burns quota Subagents are typically used for fetch-and-carry work - looking something up, running a search, summarizing a file - that doesn't need the main model's full capability. Running that routine work on the same premium tier as the main agent costs the premium rate for work a cheaper model would do just as well. ## How to fix it Configure your main agent to delegate to a cheaper subagent model rather than inheriting the parent's tier by default. The exact mechanism depends on your agent - some let you pin a specific model per subagent type, others inherit the main model unless told otherwise - so check your agent's own subagent or delegation settings for how to set a lower tier there. ## When it shows "not assessed" If your agent doesn't track subagent relationships and which model each subagent ran on, antiburn can't assess this at all - it needs both the delegation link and the delegated model's identity. For agents that do track this, a session still shows not assessed rather than clean when subagents were spawned but their model identity wasn't recorded, or when a delegated or parent model belongs to a family antiburn hasn't reviewed a premium tier for yet - an unreviewed family can't prove non-premium, so it can't support a clean result either. Partial subagent evidence blocks a clean result the same way, even with no finding observed. See [hygiene score and findings](/docs/concepts/hygiene-score-and-findings/) for how this fits into your overall score.
Overpowered subagents
antiburn flags sessions where a premium main model delegates work to subagents running on that same premium tier.
What it detects
When a session’s main-loop model spawns subagents, antiburn checks the tier of both: the dominant main-loop model (the model doing most of the main-thread turns) and every model the session delegated work to. A finding fires when the main model is on a reviewed premium tier - Claude’s opus, fable, and mythos models, or OpenAI’s top-tier gpt-6-astra and gpt-5.5/gpt-5.6 variants, or a Gemini pro model - and at least one delegated model is also premium.
Premium status is judged per model family, using a maintainer-reviewed list, not a heuristic - a family antiburn hasn’t reviewed (or an unrecognized model) can never prove premium or non-premium, so it never produces a false finding.
Why it burns quota
Subagents are typically used for fetch-and-carry work - looking something up, running a search, summarizing a file - that doesn’t need the main model’s full capability. Running that routine work on the same premium tier as the main agent costs the premium rate for work a cheaper model would do just as well.
How to fix it
Configure your main agent to delegate to a cheaper subagent model rather than inheriting the parent’s tier by default. The exact mechanism depends on your agent - some let you pin a specific model per subagent type, others inherit the main model unless told otherwise - so check your agent’s own subagent or delegation settings for how to set a lower tier there.
When it shows “not assessed”
If your agent doesn’t track subagent relationships and which model each subagent ran on, antiburn can’t assess this at all - it needs both the delegation link and the delegated model’s identity.
For agents that do track this, a session still shows not assessed rather than clean when subagents were spawned but their model identity wasn’t recorded, or when a delegated or parent model belongs to a family antiburn hasn’t reviewed a premium tier for yet - an unreviewed family can’t prove non-premium, so it can’t support a clean result either. Partial subagent evidence blocks a clean result the same way, even with no finding observed.
See hygiene score and findings for how this fits into your overall score.