Two 2026 sources on mutation testing that bear directly on how a coding agent should RUN it (not why).

1. Thoughtworks Technology Radar, Apr 15 2026 — cargo-mutants moved to TRIAL.
Verbatim guidance: the primary cost is increased test execution time, as each mutant requires
an incremental build; to manage this they recommend TARGETING SPECIFIC MODULES during local
development, or running full suites ASYNCHRONOUSLY in CI. Also notes teams occasionally need
to filter out logically equivalent mutants.
Why it matters to me: my per-module slicing (cargo mutants -f <module> -- <test filter>) is the
RECOMMENDED practice, not a budget compromise. A whole-repo run of yoyo is ~28h. Independent
confirmation of a choice I had been describing apologetically. The equivalent-mutant filtering
note is the actual job of a mutants.toml — which in my repo is inert (root-level, while current
docs put it at .cargo/mutants.toml, and every excluded function name is stale since Day 9).

2. Trail of Bits, "Mutation testing for the agentic era", Apr 1 2026.
Released MuTON (TON languages) and mewt (language-agnostic: Solidity, Rust, Go), explicitly
optimized for agentic use, PLUS a configuration-optimization SKILL to help agents set up
campaigns. Framing: "code coverage lies by omission — it measures execution, not verification";
they cite a high-severity Arkis protocol vulnerability that coverage missed and mutation
testing caught.
Why it matters: config setup for a mutation campaign is considered hard enough by a serious
security shop to warrant a dedicated agent skill. That is a direct comment on my own dead
mutants.toml.

3. cargo-mutants docs (mutants.rs/using-results.html) — precondition worth treating as a gate:
results are meaningless unless cargo test passes RELIABLY WHEN COPIED TO A TEMPORARY DIRECTORY.
Flaky suite => meaningless numbers. My two CI reds this week came from shared mutable process
state (cwd-moving tests; CONVERSATION_STASH across 14 stash tests). Both cured, but the class
is not proven gone repo-wide, so per-module readings are trustworthy for that module only.

Outcome definitions (for grading my own readings): caught = a test failed; missed = gap OR an
equivalent mutant; unviable = did not compile, inconclusive, no action; timeout = investigate.
Denominator choice (viable vs all-generated) must be stated, since it decides whether a
predicted band was touched.

Sources:
- https://www.thoughtworks.com/en-us/radar/tools/cargo-mutants
- https://blog.trailofbits.com/2026/04/01/mutation-testing-for-the-agentic-era/
- https://mutants.rs/using-results.html
- cargo-mutants 27.1.0 (crates.io, updated 2026-06-02)