Muse Code Needs a Six-Agent Trial Budget • The Weighted Average

Meta’s Muse Code beta deserves a bounded trial, not a wholesale migration: one sample six-agent standard-tier budget is $13.88, enforced by the operator rather than implied by Meta’s context window. The harness isolates each writing agent in a Git worktree, but Meta’s own six-child demonstration ran only 4 of 6 children concurrently and published no integration bill or final merge-success rate.

Six branches, four lanes, one real bill

The useful product is the operating envelope around the model. In Meta’s launch guide, each write-capable child receives a detached-HEAD worktree based on the parent commit, records its workspace lifecycle, and commits on its own branch. That keeps concurrent edits out of the parent working copy. It does not make six individually plausible changes compose into a correct product.

Meta’s Bastion Breaker fixture makes the distinction concrete. The public sample repository divides a broken game into six tasks, from one repair to five features. Six children were admitted, yet the host allowed 4 of 6 to run at once; the last two queued. One child’s suite passed inside its branch while the parent still failed. That proves workspace isolation. It does not prove the six branches merged cleanly, passed together, or beat a single-agent workflow on cost.

The sticker price lets an operator set a stop-loss before enthusiasm spends the budget. The official Muse Code pricing table charges standard usage at $1.25 per million input tokens and $4.25 per million output tokens; cached input costs $0.15 per million. At a deliberately generous ceiling of one million uncached input tokens plus 250,000 output tokens for each of six children, the arithmetic is 6 × ($1.25 + 0.25 × $4.25) = $13.875, or $13.88.

That is a sample budget, not an estimate of Meta’s demo or a natural product ceiling. The 250,000 output-token allowance is an operator-selected threshold; parent usage is excluded, cached discounts are ignored, and multiple calls can push cumulative session input beyond one context window unless the team enforces the cap externally. Meta’s Muse Spark model page says the model accepts a 1-million-token context window; it supplies no independent evidence that retrieval quality remains constant across that entire window. Reasoning tokens are billed as output, so a higher effort setting can consume the budget without producing more mergeable code.

Treat validated integration as the unit of value. Pick six independent, test-backed tasks, pin a base commit, and require the complete suite after merging each branch. This extends the lesson from GitHub’s stacked-PR response to oversized agent changes: branch geometry can make review legible, but it cannot turn local correctness into system correctness.

The cheapest tier sends the dearest signal

Muse’s tiers hide a governance decision inside a model selector. Contributor costs $0.10 per million input tokens and $0.20 per million output tokens, but Meta labels its data as usable to improve products. Standard is marked “not used to improve our products.” That is not the same promise as zero retention. The launch guide says Meta is only beginning to accept zero-data-retention requests through its sales channel, so regulated teams need written retention, deletion, access-control, and incident terms before proprietary code enters the beta.

The local event log helps, with caveats. Meta Research says Muse appends model calls, tool runs, approvals, edits, child spawns, steering, and cancellations to per-session JSONL. That makes crash recovery and execution review unusually inspectable. It also creates a sensitive artifact that can contain prompts, code, outputs, and secrets. “Replayable execution record” is accurate; “fully auditable” is not. Teams must set file permissions, retention, redaction, and deletion policy around the log.

The same research post describes a kernel-optimization case study spanning 1,000+ tool calls over up to 24 hours. Those were tool calls, not necessarily model API calls, and Meta gives no token bill, intervention count, repeated-run variance, or independent replication. Long duration demonstrates persistence and restart value. It does not establish cost-effective autonomy.

Independent proof is thin. TechCrunch’s coverage concerns the preceding Muse Spark 1.1 and competitive context, not a hands-on validation of the new harness. That absence argues for a bake-off: at least 20 representative tasks against the incumbent, recording first-pass success, jointly passing changes, merge conflicts, wall time, token cost, review minutes, recovery behavior, and security findings.

Who switches now? Teams comfortable with terminal agents and branch review can trial Muse on separable, non-sensitive work. Public-code experiments may use Contributor; proprietary-code teams should use standard only after policy review. The cost is tokens plus integration, log handling, and human verification. The conclusion breaks if the nominal context retrieves poorly, beta failures dominate, branches collide semantically, or legal terms lag operational needs.

The verdict reverses from “trial” to “switch” only if that same-repository bake-off lowers cost per validated task without reducing pass rate or weakening data policy. Until then, the right sequence is simple: mirror a non-sensitive repository, set the $13.88 ceiling, merge one branch at a time, run the full suite, inspect the event stream, and stop paying when activity stops becoming accepted code. The broader economics of CLI agents still apply: cheap tokens are expensive when tool churn and review absorb the savings. And the defensive lesson from AISI’s evaluation-governance incident is germane even outside security testing: autonomy requires an explicit budget, bounded permissions, observable action, and a human stop authority.