AgentEval 1.0.0-rc.1

Published: AgentEval 1.0.0-rc.1 and AgentEval.Cli 1.0.0-rc.1 are available on NuGet. Verified signed packages and SHA256 checksums are attached below.

Release candidate: the stable compatibility promise starts at 1.0.0. Please test this candidate against your agents and existing baselines before upgrading production workloads. See the disclaimer.

AgentEval 1.0.0-rc.1 is the first release candidate for 1.0: honest measurements, causal workflow execution, and verifiable evaluation records for .NET agents.

Install

dotnet add package AgentEval --version 1.0.0-rc.1
dotnet tool install --global AgentEval.Cli --version 1.0.0-rc.1

For an existing CLI installation, use dotnet tool update --global AgentEval.Cli --version 1.0.0-rc.1. The library supports .NET 8, 9 and 10; the CLI targets .NET 8 and 10.

What's included

• Honest measurement state. Scores, cost, timing and usage distinguish measured zero from unavailable or incomplete evidence. Reports, comparisons, assertions and exports retain those distinctions.
• Causal workflow evidence. Logical activations retain their identity through nested workflows, parallel branches, joins and replay. Producer-owned timing supports per-step budgets; missing timing cannot satisfy a limit.
• AEF by default. CLI runs write AEF records, with verified readers for Mission Control trees, timelines, compliance, history and downloads. Legacy run folders remain readable through conversion throughout 1.x, with imported custody and missing evidence disclosed.
• Embeddings and provider integration. Three embedding metrics are available through the CLI with configured instrument identity and observed per-case usage. Bounded reasoning negotiation retains explicit settings and readiness accounting.
• Stronger consumer boundaries. Expanded API compatibility guards, secure Memory report rendering, complete CLI packaging and installed Mission Control browser checks accompany the migration guidance.

Upgrade from 0.43

Read the migration guide before updating the library and tool together.

• Nullable scores, rates, costs and timings require consumers to handle unavailable evidence explicitly. Unmeasured runs remain part of reliability accounting and can leave a verdict undecided.
• Workflow counts now reflect activations or agent turns. Use the disposable root/child timing owner for producer spans; unattributed gaps remain honest gaps.
• Mock model evaluations and incomplete requested checks use exit 11. Known failures retain their failure exit codes.
• CLI persistence defaults to AEF. eval --output-dir selects an AEF destination; explicit --aef requires a destination. Capture remains off by default, and single-file reports remain available.
• CLI/sample reasoning defaults and judge budgets can change baseline comparability. Review provider configuration and recalibrate thresholds for your models.

Existing stable, Beta and Experimental boundaries remain explicit in the maturity policy. The library remains MIT; the CLI carries bundled AEF content under MIT AND Apache-2.0.

Validation and limits

PR #304 passed all 17 checks across seven workflows, including full Linux/Windows suites, AEF conformance and DCO, CodeQL, browser security regressions, documentation navigation and offline samples. The merged commit has the exact reviewed source tree. Independent whole-change review retained no findings; changed production coverage was 90.18% lines / 80.54% branches.

The successful publication workflow built the unchanged RC tag, passed 61,031 tests / zero failures / 13 skips, validated package versions and legal content, opened all 12 Mission Control routes from the installed tool, verified server shutdown, and passed 100 embedded corpus checks across three frameworks before publishing both packages and symbol packages.

Both public NuGet packages were downloaded and their repository signatures verified. Every original payload member matches the successful pipeline artifact. A fresh library consumer loaded all 14 embedded modules plus the umbrella and resolved DI services on .NET 8, 9 and 10. A fresh CLI install reports 1.0.0-rc.1 and the accepted source commit. The installed archives and all 554 extracted files under lib/ and tools/ match the signed public distributions. The workflow correction preserves the existing v1.0.0-rc.1 tag and source.

A valid AEF seal establishes record integrity, not judge correctness. Live ASSERT/Copilot Studio round-trip compatibility remains unverified. In the bounded provider smoke test, the tested model failed the OWASP assessment; this release makes no model-safety PASS claim. Known analyzer/XML warnings remain documented.

Feedback and the road to 1.0

Please report regressions through GitHub issues, including the package version, runtime and a minimal reproduction without credentials or private evaluation data. We plan about two weeks of candidate feedback before stable 1.0.0 and the aef-1.0 format tag. Another candidate will be considered if feedback requires a breaking change.

Full changelog · Changes since 0.43.0-beta