- Category
- AI Agents
- Rank
- No. 2323Tools index
Previous survey · No. 2158 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Software Testing & Quality
- GitHub
- 2 stars
- Date
About
Untyped model-checks the protocol around an AI agent harness against a TLA+ spec, both in the design and in recorded runs. Because runtimes retry, a tool call can be dropped, delivered twice or run with a lost acknowledgement, and the check tests whether the harness still gives at most one effect per step, no destructive effect without approval and no more calls than budgeted.
At a glance
Untyped uses a TLA+ model to check how an agent harness handles retries, approvals and execution budgets. It is worth reading if a lost acknowledgment could cause a tool call with side effects to run twice.
Good fit
Agent runtime developers who can map their own protocol or traces onto a small formal model.
Consider something else when
Use integration tests first if you need evidence about a specific deployed system and have not modeled its behavior.
Start here
Read the duplicate-delivery example and compare the correct and buggy models. Then map one recorded workflow to the assumptions and see which events the model can represent.
Cost and requirements
MIT-licensed. The documented checks need Java 11 or later and the repository’s model-checking workflow.
Limits to know
The result covers the modeled states and assumptions. Recorded traces do not establish that unobserved execution paths behave the same way.
Sources checked 2026-11-05. Source-based assessment.
Why it made the leaderboard
Untyped lets you model-check whether your agent harness actually enforces at-most-once effects, required approvals, and call budgets under retries and dropped acknowledgements, verified against both the spec and real recorded runs.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.