Safety checks for AI agents

Your agent says
it worked. Prove it.

Run agent output in a safe environment. Check it against clear rules. Fix simple errors automatically, and keep a record of every result.

Try the live playgroundSee the evidence
Open-source core Local-first No LLM inside the checker
run_01J9K4 · postgres-16 live
Verification runCustomer revenue by region
passed
Iteration 012 rules failed
SQL parsesread-only SELECT
DialectYEAR() is not valid
Result oraclerow mismatch
Iteration 028 / 8 rules
Dialectfixed with strftime()
Result oracle12 rows exact
Query plan34ms · index used
Stop reasonall checks passed
Total tokens668
Sandbox1.2s
No production access
Verdict attached
900+measured model runs
0LLM calls in the checker
3.8×lower mean cost vs. LLM judge
100%passed outputs checked by code
Check work before it ships

One run. Three layers of proof.

The loop controls retries. The rubric defines success. The sandbox shows how the result behaves before production.

01

Controller

Retry only failed checks. Spot repeated errors. Keep the best result. Stop when the time or token limit is reached.

  • Stops repeated failure loops
  • Safe fixes without another model
  • Focused review when stuck
02
The intelligence

Rubric

Tested rule packs for work that matters — from SQL and invoices to regulated customer letters.

  • Exact, limit, and source checks
  • Golden sets and mutation batteries
  • Rules by location and effective date
03

Environment

Safe SQL, code, and data environments that test real behavior without putting production systems at risk.

  • Fresh environment for every run
  • Safe test data and fixtures
  • Time and resource ceilings
Ready-made rule packs

Use the rules your industry already follows.

RubricLoop turns laws, schemas, policies, and business limits into checks that run on every output. People still make the final decisions.

Regulated decisions

Make sure every credit decision gives the real reason.

Check required notices, deadlines, and decision reasons before a letter reaches a customer. Send only unclear cases for review.

Reasons match the factors used in the decision
ECOA and FCRA notices are complete
The 30-day deadline is met
Rubric packfinance/adverse-action-v1
Check typeD1 · D3 · Hybrid
Safe test environments

A sandbox that tells you what passed.

Most sandboxes only tell you that a process ended. RubricLoop tells you whether the result is ready to ship and which check failed.

$rubricloop run --sandbox postgres-16
First

Relational SQL

Postgres · SQLite

Syntax, read-only access, correct results, query cost

View environment
Next

Analytical OLAP

DuckDB · ClickHouse

Memory, partitions, risky joins, result shape

View environment
Planned

Full-stack code

Python · Node · Go

Build, tests, lint, timeouts, approved packages

View environment
Planned

Data science

Pandas · NumPy

CPU, memory, files, charts, spreadsheet totals

View environment
Tested with real runs

We tested the idea before building the product.

Across SQL, invoice extraction, constrained replies, and boundary cases, every pass was checked by Python code. When the loop could not fix an output, it stopped early instead of wasting tokens.

Read the experiment summary
SQL final pass rate62.5%

+25 points over single shot, at 734 mean tokens per run.

Cost vs. naive retry−58%

Same overall final-pass rate, substantially less token spend.

Recorded abstentions144

Unclear rules were reported, never counted as passed or failed.

verify_agent.py
from rubricloop import verify

run = verify(
    agent=write_query,
    rubric="engineering/sql-safe-v1",
    sandbox="postgres-16",
    budget={"iterations": 3, "tokens": 2400},
)

if run.passed:
    deploy(run.output)
else:
    review(run.residuals)
Start simple, grow safely

Run it locally. Add managed tools when you need them.

The open-source core runs inside your stack with your model and your data. Add a tested rule pack or managed sandbox later without changing the workflow.

01
Install the open-source core

Local loops, local verifiers, local data.

02
Compose or load a rubric

Use free checks or a tested industry pack.

03
Opt into a managed sandbox

Safe execution, signed results, full history.

A clearer way to trust results

Do not ask one model to judge another.

LLM onlyPlausible

The model says it is done.

LLM + judgeSubjective

A second model reviews the first.

RubricLoopVerified

Code checks the result and saves the details.

Ready for larger teams

The developer gets an API.
The business gets a clear record.

Run locally or in your VPC Use a fixed rule-pack version Export signed results Budget and cost controls
Get started

Check your agent's work before it reaches production.

Built for teams whose agents touch databases, regulated documents, or customer actions.

Create an account Review the architecture
Start with one workflow and turn its requirements into clear checks.