The Explorer

The top rung, earned

The spectrum ended with a warning: reach for the loop last, and climb only when the world refuses enumeration. Now we're climbing — because some problems genuinely cannot be drawn as a flowchart in advance. This lesson and the three after it are archetypes: recurring shapes for spending agency well once you've established you need it. The first is the Explorer, and it is the exact mirror of the workflow we just built:

CODE AT THE FORKS          THE EXPLORER
minds in the steps,        code in the steps,
code at the forks          a mind at the forks

the path is known;         the path is DISCOVERED;
judgment fills in          judgment IS the path-choosing —
the steps                  every tool is plain, dumb,
                           deterministic code

One classification, which we'll track for every archetype: the Explorer is a single-conversation archetype. The entire investigation lives on one wall — which is both its power (every lead in view) and, as we'll see, its bill.

The Explorer's world: the terrain can only be learned by walking it. Each step's result reshapes the question — so the fork after it cannot be pre-written, not because the branches are many, but because they don't exist until the result is read.

Following the leads

⟨tool_use⟩
search_files(pattern = "4.2%")

⟨tool_result⟩
3 hits: annual-report.pdf p.7 · q3-finance-memo.docx ·
board-slides-oct.pptx

⟨tool_use⟩
read_file(path = "q3-finance-memo.docx")

⟨tool_result⟩
"…operating costs decreased 4.2% year over year (see
cost-model-v3.xlsx, 'summary' sheet)…"

⟨tool_use⟩
read_file(path = "cost-model-v3.xlsx")

⟨tool_result⟩
Sheet 'summary': YoY change = −3.1%. Sheet 'summary-adj': −4.2%
(note: "excl. one-time relocation costs, per Erik").

⟨tool_use⟩
search_files(pattern = "relocation")

This is the Explorer everywhere it appears: the auditor tracing a figure, the engineer debugging an unfamiliar codebase with grep and a test runner, the historian in an archive. The tools read the world; the mind follows the leads.

Why the tools stay dumb

It's tempting to make the Explorer's tools clever. Resist it — the archetype works because the tools are plain:

The one enhancement that fits: the retrieval rungs. Give the Explorer string search and meaning search and ask-the-document — dumb, dumber, and one-thought — and let it choose the rung per lead. The intelligence stays at the fork; the rungs are just better legs.

The map lives on the wall

Now the price, and it must be paid knowingly. The Explorer has no external flowchart — the investigation's state exists only in the conversation: what was checked, what it revealed, which leads are open, which died. The wall is the map. That means:

So the Explorer, of all the archetypes, leans hardest on the management techniques: compact into case notes when the wall bloats (leads followed → one line each; open leads → carried), file the dead trails behind addresses, and let instructive errors re-teach what the notes dropped. An investigation that never summarizes itself drowns in its own diligence.

⟨tool_use⟩
compact_conversation(max_words = 300)

⟨tool_result⟩
CASE NOTES (replacing 23 steps): Tracing −4.2% claim.
ESTABLISHED: figure originates in cost-model-v3 'summary-adj';
excludes relocation costs; adjustment requested by E. Moen
(email, 12 Oct); not disclosed in report footnotes. DEAD ENDS:
board slides (copy of memo); v1/v2 models (superseded). OPEN:
(1) was exclusion policy-compliant — audit policy §7 unread;
(2) who approved Moen's request. Archived: conv-91.

When the Explorer is the right shape

Choose this archetype when ALL three hold:
  1. the forks cannot be enumerated — each result rewrites
     the question
  2. the world is readable through plain tools — the facts
     are THERE, in files, databases, systems; they need
     finding, not explaining
  3. the task ends when something is FOUND — a source, a
     cause's location, an answer with a citation

When the task ends not with finding but with EXPLAINING —
competing stories about why, tested against evidence —
you want the next archetype.