ADR 0007 — Agent mode is opt-in, not default

Context

The tool-using agent loop (qualify_one_agent) is a strictly stronger qualification path: it can look up comp data, check GitHub activity, and read engineering blogs before ruling. It’s also 5–10× more expensive per row and 3–5× slower.

Should we make it the default?

Decision

No. Single-turn (qualify_one) is the default for qualify; agent mode is opt-in via --agent.

Rationale

The pre-filter already killed the low-signal rows. By the time qualification runs, we’re on ~5% of the raw ingestion. Most of these still don’t need agent-level scrutiny — the JD text carries enough signal for a qualified / maybe / skip decision.
Cost curves matter. At the daily batch of ~50 pre-filtered rows, single-turn is $0.50; full agent is $5+. Kelyn saves the agent budget for the top 5–10 candidates.
Agent tools depend on OpenRouter capability. Only OpenRouterQualifier implements chat_turn. Making agent default would leak provider choice into every workflow.

When to use agent mode

Top-of-list decisioning. When single-turn says “qualified” and the fit_score is >85, agent mode adds evidence-backed confidence before Kelyn spends 45 minutes drafting.
Ambiguous cases. Comp isn’t disclosed, remote status is unclear — agent can look these up.
New companies. Never seen them before, want context beyond the JD.

When NOT to use agent mode

Batch qualification of the day’s inbox.
Anything where single-turn is decisive. If the JD says “US only,” agent mode is going to reach the same conclusion at 10× cost.

Consequences

Positive: Cost control by default. Agent is a knob Kelyn turns when he needs it.
Positive: Provider portability. Single-turn works on Anthropic too.
Negative: Two code paths in the qualification runner. Mitigation: they share qualify_pending with an agent_mode: bool, so the surface area diff is small.
Negative: Users might forget agent mode exists. Documented in README.md, CLAUDE.md, and the --help text.