dispatch-strategy

Dispatching persona-respondent

Before each Agent-tool call to nemotron-personas-korea:persona-respondent, the caller picks two things — dispatch mode and concurrency — and announces both in one line so the user can override before launch.

Lane 1: dispatch mode

Mode	When
one-per-dispatch	N ≤ 3, OR interview-depth / free-text / opinion-projection (character fidelity matters; the "stay strictly in character" guarantee is strongest here).
inline batch (### Persona i=NN blocks in the dispatch prompt)	Only when per-persona prompt < 1 KB AND total batch < 5 KB. Bad fit for Korean: 3 KB CJK costs ~29 s/dispatch s_emit and collapses concurrency to ~1× regardless of cap.
file-read batch (READ_PERSONA_BATCH <abs-path> on line 1)	Default for closed-form Likert / binary / multi-choice at N ≥ 5, especially with large CJK content. ~50-byte dispatch + agent reads from disk → real concurrency. Protocol is documented in the agent file.

Validated batch size: 10 personas per agent for closed-form work.

Lane 2: concurrency (parallel subagents)

No hard cap ≤24 (measured 2026-05-30; was ~4 on older Claude Code versions — re-measure, never trust a remembered cap). Main-session Agent dispatch is launch-rate-bound, not cap-bound: the ~1.8 s/dispatch emit is the bottleneck, and wall ≈ N·1.8 s + W. Useful parallelism = Little's law L = W / 1.8 s. Dispatch in ONE continuous stream — don't batch-and-wait (barriers waste time).

Work	per-dispatch W	fan out up to (≈ W/1.8 s)
closed-form, batch-of-10	~10–15 s	~8 dispatches at once
long free-text / self-image, one-per	~45 s	~25 at once

N	subagents × personas-each
1–3	N × 1 (one-per)
4–10	1 × N closed-form; or N × 1 for free-text
11–80	⌈N/10⌉ × 10 closed-form, all launched at once (≤~8 batches run concurrently)
81–400	⌈N/10⌉ × 10, streamed (extra batches queue behind the launch emit, NOT a cap)
>1000	direct API, only if the user has opted in (see Defaults)

Beyond L, extra fan-out adds no parallelism (early agents finish before late ones launch), only context bloat. ⚠️ MAX-subscription rate limits may throttle below L at high fan-out — untested; if dispatches stall at ~20–25 concurrent, that's the real ceiling. The dynamic-workflow runtime is a SEPARATE path, hard-capped at min(16, cores−2) (=6 on an 8-core box) — don't confuse it with the no-cap Agent path.

Defaults

• Always persona-respondent for persona role-play. Don't fork to general-purpose for "closed-form factual" cases. The ~6–12% per-dispatch throughput delta (GP ~1.20 s vs persona-respondent ~1.85 s) isn't worth the per-dispatch decision tax or the silent fidelity loss when GP gets misclassified for a question that turns out to project opinion. Reserve GP for non-persona tasks (code search, planning, data inspection).
• MAX subscription via Agent over direct API. Default to Agent tool dispatches against the Claude Code MAX session. Only use AsyncAnthropic / direct API when the user explicitly opts in for the run. The OTPM math from the cross-project parallelism lesson caps Agent concurrency — it is not an automatic signal to switch to the API.

Announce format

One line, before launch:

N=10, file-read batch, 1 dispatch × 10 personas (single dispatch, no fan-out needed).

For fan-out:

N=80, file-read batch, 8 dispatches × 10 personas, all launched at once (no cap ≤24; bottleneck is the ~1.8 s/dispatch launch emit).

Skip only for trivial N=1 single dispatches.

Further reading

For the empirical model behind the concurrency and batch-size choices (NO hard cap ≤24 as of 2026-05-30 — launch-rate-bound via Little's law L = W/1.8 s; ~29 s/dispatch s_emit penalty on 3 KB CJK prompts; validated batch-of-10; dynamic-workflow runtime separately capped at cores−2), see ~/.claude/rules/lessons/claude-code-agent-tool-parallelism.md and the agent-tool-throughput skill when installed. The cap has moved 4× across versions — always re-measure, never quote from memory.