Skip to content

Augustus 0.8.0 — build and improve your own task specialist

Choose a tag to compare

View all tags
@24601 24601 released this 26 Sep 19:38
· 21 commits to main since this release
v0.8.0
94a43ad
This commit was signed with the committer’s verified signature.
ampagent Amp
SSH Key Fingerprint: EfBkhNSo6jCbYN5SnewvU6DIMCzUbgUBFH6RCBEyGjo
Verified
Learn about vigilant mode.

Augustus v0.8.0

Released 2026-09-26. Augustus now ships two skills: augustus for designing
and evaluating decision systems, and augustus-train for building and improving
an application-specific component from data to reloadable inference.

What you can ask it to do

  • Choose decision primitives, feasible base models and training methods under
    your application's error costs, latency, compute and data constraints.
  • Assemble data with provenance, adjudicate ambiguous labels, and separate
    development from confirmation without overlooking group or time leakage.
  • Fit a task specialist, export it, reload it in a fresh process, and keep
    prediction, abstention and downstream action policy distinct.
  • Run bounded improvements to data, models or programs. Compare with the
    incumbent on independent evidence; retain rules or no model when they win.

The trainer includes a runnable CPU classifier recipe. It guides an agent using
your application's tools; it is not a hosted trainer or a pretrained model.
Provider-output training still requires permission under the provider's terms.

What was executed

Fresh agents completed three journeys, then the coordinator replayed them:

  • Binary classification: three real CPU fits and bounded policy selection;
    group-weighted cost 0.023857 versus 0.107356 for always-ham on a public SMS
    proxy. All 1,025 fresh-process inference rows matched.
  • Multiclass routing with review: 150 intents plus out-of-scope examples,
    three fits and 72 policies; cost 0.117966 versus 0.3 for always-review.
    All 5,477 fresh-process rows matched.
  • No training: exact-rule routing with six correction, failure and
    fresh-process tests, plus seven separately reviewed advisory responses.

See the replayable acceptance audit
for source identities, dependencies, recipes and limitations. Repository checks
passed 190 tests plus numerical self-tests. Installation checks covered both
skills, not implicit activation in a crowded skill inventory.

These are public-proxy and fixture results, not production effectiveness or
proof that the skill beats an unassisted agent. The multiclass corpus has
near-duplicate and unknown-family dependence; its statistical bound is not clean
population evidence. The binary refit matched behavior, not serialized bytes.
This is not a recipe for reproducing a general instruction-conditioned Jev.

Upgrade

For Claude Code, remove an older pinned marketplace, then reinstall:

claude plugin marketplace remove augustus
claude plugin marketplace add 24601/Augustus@v0.8.0
claude plugin install augustus@augustus

For Skills CLI agents, install both skills:

npx skills add https://github.com/24601/Augustus/tree/v0.8.0 --skill augustus augustus-train

For manual installs, copy both directories under .agents/skills/, including
their references and scripts. The trainer uses helpers from its companion.
No existing entry point was removed. No model download, paid call or deployment
is performed by installation.

Assets 2
Loading