Ultracode vs Normal Mode in Claude Code
Normal mode works turn by turn; Ultracode works by fan-out. In a normal session, Claude spawns a subagent here and there and results flow back into its context. With Ultracode on, Claude writes a script that runs many parallel subagents with fresh context windows and returns only the final answer. You trade tokens and usage limits for breadth and cross-checking.
Side by side
| Normal mode | Ultracode | |
|---|---|---|
| How work happens | Turn by turn; occasional single subagents | Claude writes a workflow script that fans out parallel subagents |
| Context | Everything routes through Claude's main context | Subagents get fresh contexts; only distilled results return |
| Parallelism | Largely serial | Up to 16 concurrent agents (to 256 configured), 1,000 per run |
| Token burn | Predictable, per turn | Meaningfully higher; limits drain much faster |
| Mid-run input | You can steer any time | Workflows can't pause to ask you anything |
| Cross-checking | One agent's word is the word | Agents verify each other's output |
| Overhead on small tasks | None | Significant — orchestration has a fixed startup cost |
What the fan-out actually buys
The honest framing is context economics, not smarter thinking. A workflow doesn't make each individual reasoning step better — that's what effort levels do. What it buys:
- Breadth: a hundred files read by a hundred agents without any single context window melting
- Verification: one agent's proposal gets checked by another — published teardowns show workflows catching wrong technical proposals that single-pass reviews accepted
- Isolation: exploratory dead ends pollute a subagent's throwaway context, not yours
The bill for that on one real pre-check run: 471,480 tokens, 5 agents, 84 tool calls, ~4 minutes.
Task-by-task verdict
| Task | Winner | Why |
|---|---|---|
| Fix a bug in one file | Normal | Orchestration overhead dwarfs the task |
| Codebase-wide security audit | Ultracode | Breadth + cross-checking is the whole point |
| Large migration / refactor | Ultracode | Parallel per-file agents, staged pipelines |
| Quick question about the code | Normal | One lookup, no fan-out needed |
| Research with conflicting sources | Ultracode | Claims get verified by independent agents |
| Iterative UI tweaking | Normal | You need to steer between steps |
| Synthesizing a long feedback thread | Ultracode | Embarrassingly parallel reading, one synthesis agent |
| Writing a PRD or docs | Normal | Writing quality comes from depth, not breadth |
Practical middle ground
- Stay in normal mode by default; invoke the
ultracodekeyword on the occasional heavy task - If you flip the session toggle on, keep the size guideline at
smallormediumso routine tasks don't fan out wildly - Watch per-agent token usage in
/workflows— it's the fastest way to see whether orchestration is paying for itself - Pair the dials deliberately: high effort + no Ultracode for depth-bound work, moderate effort + Ultracode for breadth-bound work. More in effort levels.