Cost observability for agents · Made in San Francisco

Your agents are burning tokens right now.

Zolks shows you which ones. In real time, with the context to actually fix it.

Get on the waitlistRead the docs →

By 2030, agents will outnumber human workers online.

Every pulse on this globe is an agent making a request somewhere in the world. Right now, teams are spending hundreds of thousands of dollars a month on tokens they can't attribute, retries they can't see, and loops they can't stop.

  • $7M

    average enterprise AI budget in 2026, up 6x from 2024

  • 88%

    of teams hit surprise inference bills last quarter

  • 40x

    cost difference between an efficient agent and a leaky one

Know which agent, user, or feature is burning tokens.

$ zolks top --group-by agentagent               runs   tokens    cost    trend────────────────────────────────────────────────────refund-bot          2,847  18.4M   $146.20   ↑ 312%onboarding-agent    1,203   3.1M    $24.80   ↓ 4%support-triage      8,941   7.2M    $58.10   → flat

Get paged when an agent goes off-pattern.

$ zolks anomalies▲ refund-bot · last 1h  cost-per-run: $0.47 (baseline $0.03, 14.2σ)  root cause: 18 of 22 runs made >10 tool calls            (normal: 3 calls/run)  last normal run: 47 minutes ago

Soon: stop runaway agents before they finish the month's budget.

$ zolks budget set refund-bot --daily $50 --hard-cap✓ policy active · enforced at proxy layer