The behavior layer for AI agents
You review the diff. Nobody reviews the behavior.
Maida shows how your agent behaves, and fails the pull request check when that behavior changes in ways you didn't approve.
02 The incident
Same answer.
Different behavior.
Your agent still returns the right answer, but now it calls 3x the tools. Your output checks pass, even though the way your agent gets there has changed.
03 The behavior
How your agent actually behaves.
Which tools it calls and in what order. How many calls and steps it takes, where it retries, where it risks a loop. How latency and cost move with all of it. That is the behavior you review, and the behavior every later change is held to.
| Signal | Baseline | Candidate | Shape change |
|---|---|---|---|
| Tool sequence | lookup -> answer |
lookup x4 -> CRM -> answer |
new path |
| Tool calls | 2 | 5 | +150% |
| Steps | 4 | 8 | 2x |
| Retries | 0 | 2 | appeared |
| Loop risk | none | detected | flagged |
| Latency | 1.2 s | 3.8 s | +217% |
| Cost | $0.006 | $0.019 | +217% |
Accept change on purpose. You decide which behavior changes are acceptable. Maida compares each run with the behavior you approved, and nothing else.
04 The verdict
The verdict on your pull request.
- PASS
- The observed behavior meets your approved checks.
- FAIL
- The observed behavior violates your approved checks.
- INCONCLUSIVE
- There is not enough evidence to decide; this is never an approval.
05 Runs where your code lives
Runs on your machine or CI runner. No Maida cloud account required.
Maida's evaluation runs locally and does not upload raw traces to Maida. Your agent's own model and tool calls are a separate flow Maida does not create or remove. Anything you choose to enable states where it sends data and what it sends.
Inspect it. Open the local execution timeline.
Archive it. Keep traces with your own CI artifacts.
Control it. Export only when you explicitly choose to.
06 Install
Install Maida.
-
01
Capture one task
Use your coding agent in your repository.
-
02
Review a few checks
Choose the observed behavior you want to protect.
-
03
Try a change
See a pass, a deliberate failure, and repair.
-
04
Add CI when ready
Verify the repository boundary on a real PR.
No clone. No API keys.
A deterministic demo goes from a healthy trace to a failed check.
$ uv tool install "maida-ai==0.6.1"
$ maida demo --regression
Open quickstart
Choose the behavior you want to protect.
A short policy example from the pinned release docs:
version: 2
trials: 3
fail_fast: true
metrics:
stop_condition_reached: {kind: invariant, require: true}
step_count:
kind: measured
direction: upper
tolerance: {relative: 0.5}
Read the policy reference
Add the check to your pull request.
For a traced Python entrypoint with a reviewed baseline:
maida run my_agent.py --baseline .maida/baselines/my_agent.json \
--policy .maida/policy.yaml --format markdown
Follow the Action setup to configure a required PR check that reads the verdict.
GitHub Action: maida-ai/maida-assert@v0.6.0
Check permission: checks: write
07 Where this goes
Delegate the work. Keep the responsibility.
Agents will do more work than people can watch. That only works if their behavior is understood well enough to be trusted in advance, not inspected after. Maida starts at the pull request, where a change can still be stopped. We are building toward a standard for how agents should behave, on any model and any harness.
-
01
The pull request gate
A change in agent behavior is caught where it can still be stopped.
-
02
Accept change on purpose
Every change in behavior is reviewed by a person and accepted or rejected, on the record.
-
03
Understand how agents decide
Maida will see enough behavior, closely enough, to predict where it breaks and how to use it well.
-
04
A standard for agent behavior
Maida will state what an agent should do, on any model and any harness, before it runs.
Protecting an agent in a setup we haven't shown? Tell us what it does. contact@maida.ai
❌ Maida verdict: fail
5 of 7 checks failed
Top behavior changes
search_kb— repeated 1 -> 5 callsescalate_to_human— new tool, not in baselinesend_reply— no longer called