Claude Code subagent ideas

Subagents are scoped, isolated workers your main Claude Code session can spin up for one focused job, each with its own context window, tools, and system prompt. Here are the jobs worth handing off; and one that puts your idle capacity to good use.

What a Claude Code subagent actually is

A subagent is a separate worker your main session spawns for a single, well-defined task. It runs with its own context window, its own tool permissions, and its own system prompt, then reports back a clean result instead of dumping raw work into your main thread. That isolation is the whole point: heavy, noisy, or parallel work stays out of the way while your primary session stays focused and readable. The Claude Code subagents docs cover the mechanics; this page is about what to actually run.

12 subagent ideas worth delegating

These are jobs that benefit from a fresh context window, tight tool scoping, or running several at once:

• Parallel research sweeps: fan out several subagents to investigate different libraries, APIs, or approaches at the same time, then compare their write-ups.
• Reading piles of documents: point one at a folder of PDFs, specs, or transcripts and get back a structured summary without bloating your main context.
• Codebase exploration: "map how auth flows through this repo" as an isolated job that returns a diagram-in-words, not 40 file reads.
• Price and option comparisons: gather vendors, plans, or config choices into one clean table.
• Test runs and triage: run the suite, then summarize only the failures and likely causes.
• Log and error analysis: chew through long logs and surface the pattern, not the noise.
• Dependency and changelog review: read release notes for an upgrade and flag breaking changes.
• Migration scouting: survey every call site before you commit to a refactor.
• Documentation drafting: generate a first-pass README or API doc from the code.
• Data extraction: pull structured fields out of messy files into JSON or CSV.
• Competitive or market scans: collect and organize findings from many sources.
• Scheduled background research: a heavy-context job you run on a cadence rather than on demand (more on that below).

What makes a good subagent job

Not everything should be delegated. The tasks that pay off share a few traits:

• Scoped: one clear objective with a definable "done," so the subagent knows when to stop and what to return.
• Context-heavy: it reads a lot but reports a little, keeping the bulk out of your main session.
• Parallelizable: independent pieces that several workers can run at once.
• Tool-bounded: it needs only a narrow set of permissions, so you can scope it tightly and safely.

If a task is tiny, conversational, or needs constant back-and-forth, keep it in the main thread; the handoff overhead is not worth it.

The "do good while you sleep" subagent

Here is a scheduled, context-heavy research job that fits the subagent pattern perfectly and does something useful with capacity you would otherwise waste: Tokens for Good. It researches and vets nonprofits on a cadence, the same way you would set up any recurring background task.

You add the npm package and remote MCP server, then use /schedule to run it set-and-forget; daily, a few times a day, or weekly. Each run claims an organization, researches it, and submits a report. Because it rides your existing Claude subscription, it costs nothing extra; you are simply putting idle tokens to work. The output is not a throwaway summary: TFG uses a dual-research model with deterministic scoring and a validation pass, so results are verified before they feed the public directory at fiercephilanthropy.org/directory. See exactly how the research works or read what Tokens for Good is. To be clear, "tokens" here means AI model tokens; this is not crypto. It is the rare background job where the leftover capacity you give does real good.

Frequently asked questions