Maida provides a pre‑merge regression gate for AI agents, integrating into CI pipelines to detect behavioral changes in agents before code is merged. It records baseline runs of agents and compares new PR runs, flagging regressions such as extra steps, unexpected tool calls, loops, latency spikes, or cost increases, thereby preventing broken agent behavior from reaching production.
Funding
Funding not disclosed
Founders
Product
Problem
Engineering teams frequently modify AI agents without detecting changes to their execution behavior. Prompt tweaks can preserve final answers while altering step counts, tool calls, loops, latency, or cost, and traditional output‑only tests often miss these regressions.
Solution
Maida provides a pre‑merge behavioral regression gate for AI agents that runs inside CI pipelines. It records agent runs, captures a known‑good baseline, and compares structural properties such as step count, tool invocations, loop occurrences, latency, and cost against that baseline. When a deviation exceeds defined thresholds, Maida fails the CI job and blocks the pull request, preventing regressions from reaching production. The tool is installed via a Python package, instruments the agent entrypoint, and stores traces locally in an OpenTelemetry‑compatible format, allowing developers to review detailed timelines in a browser viewer. Policies are defined in a simple YAML file and can be integrated automatically into GitHub Actions or other CI systems.
Target Audience
Maida is aimed at software engineering teams that develop and maintain AI agents, particularly those using CI/CD workflows and frameworks such as LangChain or OpenAI Agents.
Features
- CLI and Python SDK for easy installation, instrumentation, baseline creation, and assertion checks
- Automatic detection of structural regressions: added steps, unexpected tool calls, loops, latency spikes, and cost increases
- Local‑first operation: traces are stored on the developer’s machine or CI runner, with optional export for external telemetry
- OpenTelemetry‑compatible span schema enabling integration with existing observability tools
- Configurable policy YAML to set thresholds and guardrails for step count, tool usage, loop limits, and duration
- CI integration scaffolding (`maida init`) that generates ready‑to‑use GitHub Actions workflow snippets
- Interactive timeline viewer for visual inspection of runs, showing tool calls, LLM calls, timings, and warnings