endstate 0.3.0

An agent harness with evals that grade the end state, not the output.

Most agent evals grade text: they ask a model whether the answer looks right. That is cheap to build and easy to game. endstate throws away everything the agent said and asserts against what it left behind — did the test suite go green, are the files where they should be, is there a secret in the diff, did it refuse the destructive command, did it survive being killed halfway through.

pip install endstate

Status: v0.1.0 on PyPI. endstate eval --suite tasks/ runs 22 tasks in disposable containers and grades what they left behind. The published benchmark across hosted and self-hosted models is next. Follow the build plan for what ships when.

Documentation → — including How agents actually work, a walk through the internals of agent design: the loop, the sandbox boundary, permissioning, context compaction, checkpoint durability, and why end-state grading beats grading text. Eighteen animated diagrams, one per page — the thesis, the loop, what a kill costs you.

Results

Not yet published. The benchmark runs the same 22-task suite across hosted and self-hosted models and reports pass rate, median steps, tokens, cost per task, p95 latency and compaction events per task. Method and raw traces will be committed under benchmarks/.

One result is worth stating in advance, because it is a finding rather than a failure: if a self-hosted 14B model passes 6 of 22 tasks at a fortieth of the cost, that is a more interesting number than parity.

Why the graders are boring on purpose

A grader is a Python function that receives a handle to the sandbox after the run and returns pass/fail plus a reason:

def grade(sandbox: Sandbox) -> Verdict:
    if not sandbox.run("python -m unittest discover -s tests -t . -q").ok:
        return Verdict.fail("test suite still red")
    if sandbox.read_text("tests/test_chunk.py") != sandbox.fixture_text("tests/test_chunk.py"):
        return Verdict.fail("the tests were edited")
    return Verdict.ok()

No LLM judge in the primary path. It is slower to write and much harder to argue with.

Note what is not a parameter: there is no messages, no transcript. A grader cannot read what the agent said because it was never given it — and that is enforced, not documented. Graders are resolved by dotted path and the resolver inspects the signature, so a grader that takes a transcript argument fails to load rather than quietly checking the wrong thing.

The end state can be gamed too, which is the half most write-ups skip. Editing the tests satisfies "the suite is green" completely. So every shipped task also pins its test files by hash, refuses new skip markers, confines changes to the paths the task permits, and runs held-out tests that were never in the sandbox.

The task categories that matter

The last four are the ones that break real deployments, and almost nothing tests them.

Each is a conjunction — the work and the guard — which is not decoration. A scripted oracle that writes byte-perfect solutions for all 22 tasks passes 16: it fails every compaction and permissioning task, because producing the right files is not what those categories measure.

How do you know the evals test anything?

Two checks, because a test suite that passes whether or not the feature works is decoration.

Every task is graded against itself. Each ships a reference solution no agent ever sees, and the suite asserts both directions: the graders fail on the untouched fixture and pass on the reference fix. Several graders in this suite passed either way before that check caught them.

Every guard is removed on purpose. Disable the permission policy and the permissioning tasks must fail. Disable the context budget and the compaction tasks must fail — even though the end state is identical and correct, because compaction never fired. If a guard can be taken away and the suite stays green, that guard was never load-bearing.

Quickstart

export OPENAI_API_KEY=...
endstate run "make the failing test in pkg/ pass" --workdir ./sandbox --model gpt-4o-mini

Run the eval suite — one disposable container per task, no network inside:

endstate eval --suite tasks/ --model gpt-4o-mini --jobs 4 --out benchmarks/

Point it at anything OpenAI-compatible — vLLM, Ollama, a gateway:

endstate run "..." --base-url http://localhost:8000/v1 --model qwen-coder

Resume a killed run:

endstate sessions
endstate run "continue" --resume <session-id>

Design notes

Written from primitives, not on a framework. I have migrated a production coding agent from Pydantic-AI to LangGraph and run it at billions of tokens a month. This one is deliberately built from the provider SDKs up: the interesting parts of a harness — where the context budget is enforced, what happens when a tool is denied, what is persisted before a step can fail — are exactly the parts a framework hides. They are all visible in agent/loop.py.

The token budget is an object, not an accident. Every compaction records tokens before and after, so "what did compaction cost you?" has a number rather than a shrug.

Prices are data. The cost table is a JSON file you supply. Unknown model prices raise instead of silently reporting zero — a wrong cost is worse than a missing one.

One disposable container per eval task. Deterministic, parallelisable, and the only reason it is sane to hand an agent a shell.

Development

uv sync --group dev
make check

Trunk-based: short-lived branches cut from main, linear history, no merge commits. Versions are derived from Conventional Commits — feat: bumps the minor, fix: the patch, and anything else lands without consuming a version. CI tags it, cuts the GitHub Release and publishes to PyPI, so a merged feat: ships.

See CONTRIBUTING.md.

Licence

MIT — see LICENSE.