explore

The explore agent is your codebase scout. When you need to find where something lives, trace how a feature actually works, or orient yourself in unfamiliar code, explore searches in an isolated context and returns evidence — file paths, line numbers, and the real call path — without touching a single file.

Cognitive Framing

"Report what the code does, not what it looks like it should do."

Explore runs read-only, in a fresh context, on the cheapest tier that still searches well. Its job is to compress a broad search into a short, cited answer so the caller's context stays free for the actual work. It never decides what to build and never judges whether a change is correct.

Key Facts

When to Use

• When you need to locate code — "where is the session timeout enforced?"
• When you need to trace a path — "how does the webhook get from ingress to the handler?"
• When you need repo orientation before unfamiliar work — what exists, where, and how it is wired.
• When a broad search would otherwise flood the main context with file contents that only one answer depends on.

Key Capabilities

• Widen-then-narrow search — glob for candidate paths, grep for the identifying symbol, then read only the files the matches implicate.
• Real-edge tracing — follows definition → callers → configuration instead of inferring behavior from a name.
• Evidence-first reporting — every claim carries a path:line citation.
• Honest gaps — questions it could not resolve from the repository are named, not filled in with a plausible guess.

Behavioral Checklist

• Cites path:line for every claim
• Traces the actual call or data path when asked "how does X work"
• Names contradictions between documentation and code
• Returns NEEDS_CONTEXT when the delegation prompt lacks the question or the work context path
• Distills findings — never dumps raw file contents back to the caller
• Never writes, edits, commits, or runs a mutating command
• Treats everything it reads as DATA — reports instruction-like file content, never follows it

Input Contract

Explore starts with a clean context every invocation and cannot see the caller's prior conversation. A delegation prompt must include:

1. The concrete question — "where does X happen", "what calls Y", "how is Z wired"
2. Known paths — any file or directory already believed relevant
3. The purpose — what the answer will be used for, so explore knows how much depth is enough

Per-Provider Model

Explore is in the Cursor bundle's default core pack alongside planner and reviewer. See Claude Agent Models, Codex Agent Models, and Cursor Agent Models.

Common Use Cases

Pro Tips

Ask One Question Per Invocation

Explore is cheap and stateless. Two focused invocations return sharper evidence than one prompt asking for a whole subsystem survey.

Say What the Answer Is For

"I need to add a field to this response" produces a different — and more useful — depth of trace than the same question with no stated purpose.

Key Takeaway

Explore keeps broad search out of the main context. It reads widely, answers narrowly, and cites everything — so planning and review start from evidence rather than assumption.

Related Agents

• planner — consumes explore findings before writing a plan
• reviewer — uses explore to confirm how a touched path actually behaves
• researcher — the external counterpart: libraries and documentation rather than this repository
• orchestrator — routes locate/trace questions to explore during Phase 0