PUBLIC BENCHMARKS
Strategy Bench
Repeated match-play across planning, deception, cooperation, and risk tolerance produces a behavioral fingerprint, not a leaderboard. Hosted on ClashAI as a public benchmark.
Verify every agent change before it ships.
Turn agent traces into deterministic test environments. Re-run every model, prompt, and tool change before production.
Private beta · For teams shipping autonomous, tool-heavy agents in production.
Private by design
Your traces, replay environments, and signed reports stay controlled from ingestion through verdict.
Model & stack agnostic
Works above your tracing, model, prompt, tool, and MCP layers instead of replacing them.
Change impact reports
Every model, prompt, or tool change resolves into quality, cost, latency, and reliability deltas.
Auditable & reproducible
Pinned seeds, suite versions, budgets, and trajectory replays make the verdict defensible.
HOW IT WORKSAgents fail in production.
Agents fail in production.
Catch them before.
Works with your current stack.
Trace in.
Your traces become a deterministic replay environment.
Re-execute.
Swap a model, prompt, or tool. One variable changes.
Verdict out.
Deploy / review / block. Signed in 72 hours.
CHANGE IMPACT REPORT
See what changed before the change reaches production.
- Compare every model, prompt, tool, or MCP change against the pinned baseline.
- Same scenarios. Same seeds. Like-for-like.
- Quality, cost, latency, and reliability deltas, side by side.
- One verdict per change: deploy, review, or block.
RUN 0488BASELINE VS CHALLENGER
QUALITY+4.2%
COST-9.1%
LATENCY+0.22s
RELIABILITY99.1%
Model swapgpt-5.1 → gpt-5.2
-18%+220msReviewPrompt editrefund policy v12
+3%-90msDeployTool changeMCP billing server
+11%+480msBlockBuilt on open research and open infrastructure.
Taso is the commercial product. The substrate underneath it is open and credible.
RESEARCH
Deterministic Reward Hacking
Identifies agents optimizing the metric instead of doing the work, the failure mode observability misses. NeurIPS 2026 submission.
OPEN SOURCE
Environments, Adapters, and Scorers
MIT-licensed environments, adapters, and inspect_ai-compatible scorers behind every Taso report.
All three are built on the same substrate: pinned fairness, reproducible artifacts, deterministic re-execution. The methodology behind every Taso report.
Currently in private beta.
Request access to Taso Labs.
We're working with a small set of design partners shipping agents in production. Two-week pilots, fixed fee, one workflow, one signed report. Earlier than that? If you're building agents and the validation gap is on your mind, tell us what you're working on. We read everything.