free-handoff: Cost Engineering for Multi-Model Agent Orchestration

1 · Why a harnessthe problem

Much research work is wait-dominated: running SAS/CIAO reductions, waiting for spectral fits to converge, waiting on Virgo-server pipelines. Handing such tasks to Opus/Sonnet-class agents spends money on "the model watching a progress bar" — every turn replays the full context, cache reads/writes dominate, and a few hours of polling costs $10–25 per task.

The free-handoff harness idea: a dispatcher (the main Claude agent) writes each task as a brief and hands it to an executor arm drawn from a tiered model pool. Every arm runs in an isolated sandbox and must deliver reproducible checks (runnable commands + SHA256). Critical tasks run as paired arms (same brief, two models from different families), with the dispatcher independently recomputing and adjudicating. Free models carry the waiting; paid models are spent only where judgement matters.

2 · Four tiers and pricingtiers & pricing

Four tiers (finalized 2026-10-16), ordered by API-equivalent price:

Prices (USD/M tokens, checked 2026-10-11): Sonnet 5.5 input 2.00 / cache-read 0.10 / cache-write 2.50 / output 10.00; luna input 0.10 / cached 0.01 / output 0.50; Haiku ≤100K prompt 0.10/0.01/0.50, >100K 0.50/0.05/2.50; Opus 4/20. Subscriptions are paid, so API prices are a yardstick, not an invoice.

3 · Paired arms: cost vs capabilitypaired arms

Same brief, two arms, independent recomputation by the dispatcher — the mechanism that makes free-handoff's comparisons meaningful. Below are the most representative pairs from the ledger (capability 0–3: 3 = correct at first delivery).

Sonnet vs sol (both normal tier, different vendors)

Sonnet vs free arms (the widest cost gap)

Sonnet vs luna (normal vs cheap)

Sonnet vs Opus (normal vs expert)

4 · Cost-control resultsresults

The core result: move wait-dominated work (SAS reduction, batch fitting, long polling) from normal/expert down to free/cheap — a 10–100× cost reduction at equal capability. Judgement-dense work stays on normal; expert appears only as a control arm (once per 5 normal tasks, adaptive ratio 3:1 → 7:1).

5 · Routing lessons: which task to whomrouting

The ledger answers "who gets what" from several hundred runs. Conclusions as of 2026-10-17:

6 · Guardrails learned from failuresguardrails

The ledger's Routing notes accumulate 60+ dated lessons. The structural ones:

Cost side

Rule 1  Wait-dominated long tasks: the brief says "launch-and-exit" — the executor starts a detached worker,
       writes the report skeleton, and exits; a watcher (or the heartbeat) resumes it later for the report.
       Avoids full-context polling.
Rule 2  Do not resume a long thread for a costly arm's stage 2: every turn replays the entire context.
       sol's stage-2 resume cost $26.08 (178M input). Use "fresh session + compact hand-off" instead,
       unless the stage truly depends on in-context state.
Rule 3  Quota gate first: no dispatch unless quota-gate claude/codex passes; a 5h window already reset counts
       as full (--rolled-5h-ok), but a live 5h reading below 30% is still NO-GO — it would cut off the
       running subagent.
Rule 4  Never price a run from the Agent tool's printed token total (e.g. 235,146) — it is not a billing
       measure; recompute exactly from the transcript events with cost.py.

Quality side

Rule 5  Dual briefs must say in words "do not open the other arm's REPORT / out_* files" — otherwise the
       second arm read the first's report, its agreement is copied, and the run is not scorable.
Rule 6  Give the CPU gate as an exact snippet (busy = Δ(total−idle−iowait)/Δtotal); never let the
       executor write its own — two self-written gates counted idle as busy and waited 1.5 h for nothing.
Rule 7  Brief defects dominate errors: two thirds of executor "mistakes" were dispatcher brief bugs.
       Validate the brief's paths, target existence, and naming rules on the repo before dispatch.
Rule 8  Hermes non-interactive bans: never end the turn while a background job runs; never ask
       ("shall I write the report?"); never repeat an identical failing call (observed: 283 repeats).
Rule 9  codexnim silent stalls cluster near ~200k request tokens: do not resume a ≥200k thread —
       start a new session with a short continuation prompt pointing at saved products.
Rule 10 Numerical fitting briefs must set OMP/OPENBLAS/MKL_NUM_THREADS=1 per worker — multithreaded
       BLAS once pulled load 65 on 32 cores from 4 workers.

7 · Reusable mechanismsreusable

If you take five things away:

1. Tiers + controls: a four-tier model pool; free arms carry volume, normal arms supply control arms at an adaptive ratio (3:1–7:1), expert appears once per 5 normal tasks. All controls share one brief, and the dispatcher recomputes independently.
2. A cost ledger: every run records wall time, tokens, API-equivalent cost, capability 0–3, errors, corrections, and whether it was used. One week of data answers "who gets what".
3. Capability scoring, not vibes: 0 unusable / 1 needs a redo / 2 correct after minor corrections / 3 correct at first delivery, against a dispatcher-recomputed reference.
4. The brief is the contract: paths, acceptance checks, resource caps, and bans are all written down; executors may not silently substitute, write outside bounds, or deliver anything non-reproducible (commands + SHA256).
5. Failures enter the ledger: every incident becomes a dated rule that the next brief cites. Standing red teams (≥2 free per project), findings triaged as material / minor / false alarm and recomputed.