1 threshold model
Others Require Alert Thresholds
You have to instrument, tune thresholds for each data stream, and anticipate every failure mode worth watching. Miss one, and you're blind to it.
AI SRE
The AI SRE that catches issues before you get paged
Herald detects potential issues, investigates them across code, infrastructure, and telemetry, and hands you the root cause — all before an alert fires or a customer complains.
SIGNAL-4821 / slack-ingestion 3,490 signals · 2 error groups
07:06 UTC - Anomaly detected
HTTP 500 spike on Slack event ingestion pipeline
- Anomaly detected 5s later
Outbound timeouts in
slack_bolt→aiohttpauth path - Cluster grouped 30s later
2 related error signals correlated across alert pipelines
- Investigating 10s later
- Correlating error groups across ingestion and alert pipelines
- Tracing
auth.testcall in Slack Bolt authorize hot path - Matching prior
aiohttp→slack_sdktimeout pattern
- Root cause determined 2m later
auth.testAPI call timing out during authorization, causing intermittent 500s on the ingestion hot path.Cache
AuthorizeResult· configure explicit SDK timeouts
Real predictive detection on Herald production; sensitive details redacted. See more
Trusted in Production
Why observability is broken
High Maintenance. Poor Coverage.
The metric stays below the alert threshold for most of the chart, then crosses it. Marker 1 identifies the threshold crossing. The area above the threshold after that crossing is labeled no runbook coverage, and marker 2 identifies the uncovered alert territory.
2 runbook coverage
Others Require Runbooks
You document every investigation workflow before it's needed. Maintain them as your stack evolves. When something novel breaks, there's no runbook and no investigation.
Other agents only handle failures someone already documented. Herald investigates the ones nobody saw coming. No runbook required.
THE HERALD APPROACH
Stop reacting. Start preventing.
-
Learn
Herald builds a context graph before any alerts fire — observability, codebase, CI/CD, docs, and dependencies — so it knows what normal looks like and can work to solve any problem.
context: Jira tool 65 · config read path ·
CUST-8291-X -
Detect
No thresholds to set. Herald builds a custom anomaly detection model for each data stream and surfaces validated issues before your customers notice.
validated signal: HTTP 500s rose to 18–26% over 22 minutes while other tenants stayed flat
-
Investigate
Never write another runbook. Herald evaluates multiple hypotheses simultaneously, each against the right data source, and delivers RCAs in minutes.
RCA: Schema Drift · legacy Vault keys rejected after
PR #4275
In Production
Results. Delivered Fast.
Heralds's agent onboards and adapts quickly. Gartner's 2026 AI SRE Market Guide identifies proactive incident prevention and contextual awareness as next-generation capabilities. Herald already does both.
- Results in days, not months. The Herald agent learns your stack quickly and efficiently – see your first RCA in days.
- Solves the unknown. 70%+ accuracy on novel incidents for one of the world's biggest B2B2C platforms.
- Never repeats mistakes. Herald learns from every single investigation, so it never makes the same mistake twice.
05 — The Team Behind Herald
Powered by UC Berkeley research
Herald was founded by PhDs and Professors from UC Berkeley's innovation center, RISELab, combining expertise in AI, LLMs, data systems, and scalable infrastructure.
FROM THE TEAM
View all postsRead the latest
Why You Should Definitely DIY Dev Tools with AI. Sometimes.
Engineering teams are building their own AI dev tools more than ever. After comparing notes with a lot of them, here's where DIY pays off and where it doesn't.
ClickHouse + Herald: AI DevOps intelligence on fast, cost-efficient telemetry
Herald, an AI DevOps agent in the House Mates program, puts the telemetry you keep in ClickStack to work: predicting incidents, delivering root cause in minutes, and answering questions about your systems. ClickHouse customers can get up to $25,000 in free Herald usage.
No Runbooks, No Problem: Snorkel AI Gets Day-One Results with Herald
93% accuracy on engineering questions. Incidents resolved in minutes. No runbooks. No RCA documents. No Slack history.
