Route work, control changes, recover from failures, and inspect the result.
Choose an engine · See a real proof workflow · Engineering notes
AEB builds independently usable tools for developers and researchers working with agents, software, and computational workflows. Each engine has a specific responsibility and a bounded contract. Use one on its own, or compose the public interfaces a task actually needs.
Domain-specific models and student-facing interfaces belong in separate applications — not inside the execution, routing, transaction, or verification engines.
| You need to... | Engine | Its responsibility |
|---|---|---|
| Route work under explicit constraints and recover bounded multi-step execution | GaugeMesh | Capability routing, durable Tasks/Runs, persisted state, and explicit reconciliation |
| Inspect, accept, or roll back supported coding-agent changes | AgentTX | Repository transactions, declared validation gates, and inspectable proof packs |
| Reproduce MCP failures and retain a regression | ResiliReplay | Controlled fault injection, bounded recovery evidence, replay, and generated tests |
| Evaluate explicit claims before continuing a model workflow | VerifAxis | Verifier-conditioned recurrence, stopping decisions, and evidence inspection |
| Preserve a qualitative transition in a user-provided simulation | PhaseProbe | Bounded transition discovery, replayable fixtures, and generated regression tests |
Independent release proof: ResiliReplay Action Smoke checks the released ResiliReplay Action from a separate consumer repository. It is evidence infrastructure, not another required runtime or a competing application.
Application / agent / client
|
+-- Route and execute ........ GaugeMesh
+-- Control repository edits . AgentTX
+-- Test failure and recovery ResiliReplay
+-- Evaluate explicit claims . VerifAxis
+-- Preserve transitions ..... PhaseProbe
Released artifacts ................. independent downstream checks
These are optional interfaces, not a fixed pipeline. Files, documented CLIs, and supported protocols connect components; importing every engine or installing the entire portfolio is not required. A specialized application owns its subject-matter rules and user experience.
Changing code with an agent? Start with the AgentTX workflow and proof gallery. The gallery shows declared validators rejecting a test-weakening change and the resulting transaction record. It is not a general detector of every unsafe edit.
Building an MCP integration? Use the ResiliReplay quickstart to exercise the documented failure/recovery path and inspect its artifacts.
Need restartable execution? Read GaugeMesh's public interfaces and limits and choose a supported Task or Run path.
agenttx proof --validator '["npm","test"]' -- codex exec "fix the failing test without weakening it"
Real AgentTX Proof Card: a weakened test was rejected and rolled back
The three-case proof gallery shows a declared policy accepting a supported change and rejecting a weakened test. Inspect the receipt and declared gates; this is bounded transaction evidence, not proof that every possible bad edit or external side effect is detected.
Portfolio engineering evidence is a historical record of the observations made at the recorded dates, not a current cross-project version inventory.
| Project | Releases | Usage and boundaries |
|---|---|---|
| GaugeMesh | GitHub releases | README |
| AgentTX | GitHub releases | README |
| ResiliReplay | GitHub releases | README |
| VerifAxis | GitHub releases | README |
| PhaseProbe | GitHub releases | README |
Use the chosen release's exact installation instructions, supported platform, and versioned contract. Availability of an artifact does not imply production readiness or support for an untested platform. Preview and research status remain visible in the project releases.
Execution, artifact integrity, policy acceptance, and independent numerical checks answer different questions. A successful run is not automatically a correct model or a safe design.
AgentTX is not an operating-system sandbox and cannot undo arbitrary external side effects. GaugeMesh's recovery and cancellation guarantees are bounded by its documented execution surface. Hashes and unsigned receipts are not producer authentication. VerifAxis depends on the scope and quality of its verifiers; PhaseProbe depends on the supplied observable and predicate. Exercise effectful targets only with explicit authorization.
Report a reproducible failure, unclear quickstart, or real compatibility result in the relevant repository. Share only sanitized examples — never credentials, private prompts, research data, or proprietary files.
Useful tools earn support through use. Feedback and repository stars are voluntary; they are never required to install, run, or verify a result.