I build the software around AI models: tools that expose useful capabilities, keep execution inspectable, and make results testable.
A long test log should not consume the context needed to fix the failure. Straitjacket captures tool output, returns a bounded result, and gives the agent addresses for retrieving the omitted evidence. Its edit workflow connects observed source, anchored changes, verification, and continuation evidence.
First proof: run one noisy command with ctx run, then retrieve the lines
behind the result. Compare with native tools on your own task; the published
benchmarks include cases where the extra layer loses.
Repository and quickstart · Evidence · Docs
A CLI and an MCP server should not disagree about whether a refund needs confirmation. Anvil compiles an API description and a reviewed manifest into interfaces that carry the same policy, plus mocks and evidence for checking it.
First proof: compile the payments fixture, inspect the refund policy, observe a refusal, then produce a dry-run request plan with its requirements satisfied. No upstream payment service is needed for that workflow.
Repository and quickstart · Supported inputs · Docs
Plausible documents are not enough for an evaluation. Worldloom builds related
facts, artifacts, and evaluation truth from a synthetic world, then renders
files such as spreadsheets, documents, slides, and PDFs. The repository is
named synthetic-foundry; the product and CLI are Worldloom and worldloom.
First proof: build a seeded incident corpus, render it, validate its coherence, and run the retrieval baselines. Coherent source data and successful retrieval are separate results, both available for inspection.
Repository and quickstart · Guides · Docs
These products address different parts of an agent workflow. They can be assessed independently; this is not a claim of a packaged integration between them.
| You need to… | Start with | Evaluate first |
|---|---|---|
| Recover an acting agent after interruption | Tape | Journal, replay, and effect recovery |
| Build a Rust text or live-voice agent | gemini-rs | One text request or voice session |
| Write native ADK agents with less setup code | adk-fluent | A builder and the native object it returns |
| Turn agent requirements into reviewable implementation evidence | GE Agent Factory | A local proof pack |
| Apply policy to agent, tool, and model traffic | AIPlex | The traffic plane you need |
| Convert authored HTML into PowerPoint | Pixelpitch | Export fidelity and element editability |
| Use Gemini Enterprise inside Microsoft 365 | ge-msft | The supported host and document operation |
| Explore an agent authoring interface | Ember | One playbook through the frontend prototype |
The repository directory separates original projects from forks and links older experiments without presenting every repository as a supported product. Each project's README is the starting point for its setup and limits.
Earlier work includes autonomous-driving experiments, cloud operations, Kubernetes, and application integrations. The recurring interest is making uncertain behavior observable and testable.
Useful output before architecture tours. Reviewable policy before unattended effects. Reproducible evidence before performance claims. A small first run that makes the next decision clear.