At the time of this writing, sub-agents are still pretty new. (Like, less than a week old.) So, they still have some rough edges.

• Inconsistent activation — Claude often ignores an appropriate sub-agent unless you name it explicitly. Even with a solid description in claude.md, auto-selection fires only sometimes, so “fire-and-forget” delegation isn’t reliable yet.
• State loss on rejection — If you decline an agent’s draft, main Claude spins up a fresh copy rather than letting you iterate with the same one. All of that agent-specific context disappears, forcing a ground-up re-run.
• No mid-stream dialogue — You can’t ask a running agent clarifying questions or nudge it part-way through a task. The black-box execution means you wait in silence, then accept or reject a lump-sum result.
• Opaque inner workings — Internal tool calls and partial thoughts remain hidden. When an agent chews through 50 k tokens you have no incremental visibility, which makes debugging slow or impossible.
• Token & time bloat — Each sub-agent adds another context window. Large fleets (10-15+) eat through the 200 k token budget quickly and can run for an hour+ before surfacing anything useful.
• Performance degradation with “too many chefs” — A long agent list dilutes relevance signals; the delegating model spends cycles choosing instead of doing. Overlapping responsibilities cause ping-ponging or duplicate work.
• Tool-scope confusion — Giving every agent every tool sounds convenient but introduces noise. The model sometimes picks an ill-suited tool or fails because it can’t decide which one to use.
• Short, shallow outputs — Even well-defined agents can return a single-sentence verdict (“yeah, looks good”) instead of the requested deep dive. You still need follow-up prompts to get actionable detail.
• Verbose auto-generated prompts — The /agents wizard produces sprawling system prompts. These eat context, slow reasoning, and still don’t guarantee correct activation. Most users end up trimming them manually.
• Non-deterministic variance — The same task with the same agents can yield different plans—or different no-shows—run to run. Measuring improvements is tricky because signal swings with each session.
• Activation edge cases — Agents created for highly specific events (e.g., post-file-save code-guardian) may never trigger unless the main prompt or repo state matches their narrow pattern exactly.