01 — The setting
Start with the
real workflow.
Compare a baseline and a candidate on the same tasks, with the same tools, permissions, and budget.
Start with coding agents. Extend the checks as tools, environments, and memory evolve.
Independent AI evaluation
Know what
you’re releasing.
We’re building independent evaluation to help teams iterate toward more reliable agents and safer releases.
Explore our approachReliability. Safety. Every iteration.A closer look
Emerging agent behaviors
New capabilities.
New behaviors to understand.
We’re designing evaluations for emerging behaviors across computer-use agents, multi-agent systems, connectors, memory compaction, and recursive self-improvement (RSI) systems. Each iteration brings new questions about reliability and safety.
Our approach
From a run
to a reason to ship.
Hold the conditions steady. Make the difference visible.
Start with the
02 — The behavior
Look beyond
task completion.
Can agents use computers reliably, respect connector permissions, coordinate across agents, and retain critical instructions through memory compaction?
Check task success and safety separately. Connect each finding to an inspectable trace.
03 — The next iteration
Put evidence behind
the release decision.
Learn from each iteration. Re-run failures, track regressions, and bring the evidence into the next release decision.
Keep uncertainty visible. Make the next check clear.
The evidence package
A result you
can look inside.
We’re designing the report around what a release owner needs to review.
Design & research01Release comparison
What improved. What regressed.
Under the same conditions.
Release comparison
What improved. What regressed.
Under the same conditions.
A baseline-to-candidate comparison with the task set, environment, permissions, and budget recorded alongside the findings.
02Failure evidence
A finding is the starting point.
The trace makes it inspectable.
Failure evidence
A finding is the starting point.
The trace makes it inspectable.
Failure cases linked to the actions, logs, and artifacts behind them, with reproduction steps where available.
03Coverage & uncertainty
What was tested.
What still needs an answer.
Coverage & uncertainty
What was tested.
What still needs an answer.
Untested behavior, incomplete runs, and limits on what the evidence can support remain part of the report.
04Cost & efficiency
What it took to get the result.
Across the whole run.
Cost & efficiency
What it took to get the result.
Across the whole run.
Our goal is more accurate cost evaluation: model, tool, and compute usage across agents, retries, and coordination. Compare cost per successful task alongside quality, with measured usage, pricing assumptions, and estimation gaps kept explicit.
Built for every iteration.
Our product goals
Fast.
Less setup. Shorter turnaround.
From a question to reviewable evidence.
Secure.
Isolated execution. Scoped permissions.
Controlled access to sensitive data.
Reliable.
Recoverable runs. Traceable results.
Every attempt accounted for.
Let’s start with a real question
Which release decision
needs better evidence?
We’re in the design and research stage, starting with teams shipping coding agents.
Start a conversation