Weave

Sazabi turned what used to be blind spots into issues we catch and fix before customers ever notice. The next best thing to preventing a problem entirely is knowing about it right away, and that’s exactly what Sazabi gives us.

About

Weave gives engineering leaders a clear view of how AI changes team output, code quality, and AI spend. By July 2026, Weave had analyzed more than 2 million human and AI code contributions across 20,000+ engineers at 1,000+ companies, including Robinhood, PostHog, and Telnyx. That same month, the company announced a $13.5 million Series A led by Standard Capital.

Challenge

Static Alerts Could Not Keep Up With Changing Production Patterns

Weave tracks engineering work from prompt to production, then uses those signals to route each task to the most cost-efficient AI model without compromising speed or quality. Eight engineers keep the platform running across Google Cloud Run, Postgres, and self-hosted ClickHouse, leaving limited capacity for manual log reviews or sorting through noisy alerts.

As Co-Founder and CTO, Andrew Churchill moves between hands-on engineering, technical planning, customer work, and company direction. Weave already had logs, tracing, and coding agents connected to its observability data, but no reliable way to proactively surface unexpected production issues.

“Sometimes an alert caught it,” he says. “More often, a customer reported the problem, or we noticed something off in production, which is not how you want to learn about an issue.” When a meaningful signal went unnoticed, the team had to dig through logs to identify unexpected behavior and determine whether it required action. While Google Cloud alerts covered latency and queue depth, static thresholds could not always separate normal demand from a real problem. A queue warning fired whenever waiting jobs exceeded a preset threshold, including periods when legitimate traffic accounted for the increase.

The noise compounded when AI clients calling Weave’s MCP server tried unexpected argument combinations, even when the interface documented how it should be used. Those unpredictable requests created failure modes that were difficult to anticipate with predefined rule-based monitoring. Engineers spent time sorting through noisy signals, while important classes of behavior still fell outside predefined monitoring rules.

Without a scalable observability workflow, Drew Bailey, Founding Engineer at Weave, became what Andrew called its “human observability solution.” He periodically reviewed logs, fixed issues he found, or routed them to the right owner. The safeguard was useful, but repeatedly inspecting production by hand took time away from other engineering priorities.

Andrew needed earlier detection without manually predicting every failure condition or adding more alert noise. Sazabi stood out because it could identify unexpected production behavior from Weave’s existing logs without requiring the team to define and maintain a monitor for every possible failure.

We could investigate an issue once we knew about it, but we needed a better way to learn when something was wrong. Sazabi was the first approach I saw that addressed the problem in a way that made sense for us.

Solution

Sazabi Turns Production Logs Into Actionable Signals

Weave started with a limited rollout, connecting data sources and configuring permissions. Sazabi then opened pull requests for the small code changes its agent needed. Initial setup took less than 10 minutes. Andrew reviewed the early alerts, then asked Drew to validate the signal against the production patterns he had previously monitored manually.

Once they were confident in the early signal, Andrew opened the Slack channel to the rest of the engineering team. “Everyone picked it up naturally,” he says.

Now, Sazabi analyzes Weave’s logs for unexpected behavior and posts detected issues to Slack, where all eight engineers can see them. Instead of scanning the logs for errors or unusual production patterns, Drew starts with a specific issue Sazabi has surfaced, then fixes it or tags the engineer who owns the affected service. Sazabi takes on the open-ended search while Drew keeps the judgment and ownership. When an alert reflects expected behavior, Andrew gives Sazabi the missing operating detail. The platform retains that context and avoids raising the same class of issue again. If Andrew identifies a gap or wants to adjust a workflow, the Sazabi team responds quickly and works with Weave to resolve it. “Their team is super responsive,” Andrew shares. “We can ask, ‘What’s happening here?’ or ‘How should we configure this?’ and they’re right there to help us make it work the way we want.”

The same feedback loop sharpens Weave’s job queue, where the amount of work waiting to be processed naturally rises and falls with traffic. Instead of alerting whenever that backlog crosses a preset limit, Sazabi compares it with previous traffic patterns and notifies engineers when the buildup is genuinely unusual. That reduces noise and keeps the team focused on conditions that may require investigation.

Sazabi also surfaces unexpected errors on Weave’s MCP server. When AI clients send unexpected argument combinations, it flags the resulting errors without requiring Weave engineers to configure a bespoke monitor for each one. The team receives a Slack notification when Sazabi detects the error, rather than learning about it later through Drew’s next log review or a customer report.

For issues requiring a code change, Sazabi fits into the coding-agent workflow Weave already uses. The team can hand the issue and supporting evidence to its preferred coding agent or use Sazabi’s one-click fix to prepare a pull request. An engineer reviews the change, and once approved, Weave’s continuous deployment process ships it in under ten minutes.

With Sazabi, Weave now starts with evidence-backed investigations, moving from hidden failure to a reviewable fix faster.

Sazabi gives us the alert channel we wanted. It tells us when something is actually wrong without requiring us to define every condition ourselves.

Results

20+ Engineering Hours Saved Each Week

For Weave, partnering with Sazabi makes proactive production reliability more sustainable. By reducing manual monitoring through earlier, high-signal detection, Andrew and his team can free up engineering capacity for investigation, remediation, and prevention.

• <30s to detect, 8.5 minutes to deliver full root cause

• <10 minutes for initial production setup

• 20+ engineering hours saved each week

Going forward, Andrew is continuing to move static-alert workflows to Sazabi. That gives Weave a scalable way to identify unexpected behavior as production changes, without defining and maintaining a new threshold or log-review process for every failure mode.

We had a collection of static alerts before. Now, we keep replacing more of them with Sazabi.