From DevOps ‘Heart Attacks’ to AI-Powered Diagnostics With Traversal’s AI Agents

Traversals Mission & Market Opportunity: Anish and Raj, co-founders of Traversal, are building AI agents to automate Root Cause Analysis (RCA) in DevOps and Site Reliability Engineering (SRE), aiming to replace manual "war room" troubleshooting with scalable agents.

• Current SRE workflows are analogized to a healthcare crisis where teams constantly treat "heart attacks" (high-severity incidents) and "chronic conditions" (daily alerts), leaving no time for strategic "life hacking" (infrastructure planning).
• The rise of "vibe coding" and AI-generated code in non-critical apps creates a "fast fashion" software ecosystem, whereas mission-critical sectors (payments, finance, healthcare) require reliable, maintainable systems that current AI coding tools fail to guarantee at scale.
• A "tale of two worlds" is emerging: one where AI code is discarded like fast fashion, and another where AI tools generate mission-critical code that requires complex maintenance, creating a bottleneck for human debugging due to a lack of code context.

Current State of Observability & RCA: Despite observability being the second-largest software spend (after cloud), the RCA process remains manual, fragmented, and chaotic.

• Incidents typically trigger 30–50 person Slack channels where blame is assigned and tribal knowledge is required to solve issues, relying on "10x engineers" who cannot be replicated.
• Existing tools (Datadog, Splunk, Grafana, etc.) focus on storing and visualizing MELT (Metrics, Events, Logs, Traces) data but fail to automate the complex workflow of connecting disparate data points to find the root cause.
• Traversal differentiates by automating the workflow of troubleshooting rather than just the storage of data, operating on the principle that the answer lies within the data but is too large for any single human to process.

Product Architecture & Capabilities: Traversal's agents function as LLM orchestration systems that sequentially call tools (data fetching, anomaly detection, statistical analysis) to build context and solve complex RCA tasks.

• The system relies on "inference time compute" rather than hard-coded prompts or workflows, allowing the model to dynamically decide steps based on the specific enterprise's data structure.
• An "offline phase" (5–10 hours for small customers, up to a day for large ones) is used to build a semantic dependency map of logs and code, leveraging statistical causal inference techniques derived from the founders' academic research on gene regulatory networks.
• The "online phase" uses this map and real-time data to guide the agent through a systematic search for the root cause, mimicking an SRE's logic but scaling it across thousands of microservices.
• Traversal currently achieves ~90% accuracy in identifying root causes for incidents where the answer exists in the data, resolving issues in 2–4 minutes, a significant reduction from the hours spent by human teams.
• The company rates its agents at L4 on the self-driving car scale: they can reliably identify the root cause (e.g., specific log or PR) but cannot yet autonomously execute complex, system-wide fixes (L5).
• The system operates with read-only access to customer data to avoid security friction and new data generation requirements.

Strategic Decisions & Challenges:

• The "0% Accuracy" Wake-up Call: Early versions of the system achieved 0% accuracy on large-scale enterprise incidents when using hard-coded workflows; accuracy only skyrocketed after pivoting to an architecture leveraging inference-time compute.
• Enterprise Integration: The product is agnostic to data storage, avoiding the "security hell" of bringing proprietary models to enterprises that often mandate using their existing LLM providers (e.g., OpenAI, Anthropic via Azure).
• Customer Suitability: The solution adds the most value to large enterprises with mature observability but fragmented human context, whereas it adds less value to Series A startups lacking the necessary data instrumentation.
• Future-Proofing Strategy: The company bets on the continuous improvement of reasoning models, architected their system to maximize these capabilities, and accepts that they must re-evaluate their stack every six months to stay ahead of AI advancements.

Future Outlook & Team Dynamics:

• Team Composition: Traversal's team is 90% engineers with a mix of traditional software backgrounds, ML PhDs, and infrastructure experts; the barrier to entry is lowering, making "experimental mindset" more critical than formal credentials.
• Evolution of SRE Roles: In five years, SREs will need fluency in both traditional systems and AI-specific failure modes, shifting from "ICU surgeons" to strategic planners and agents that handle the "creative fun" of infrastructure design.
• Logging & Instrumentation: The definition of a "log" will fundamentally change from human-readable text to AI-optimized data structures that explicitly connect business logic to technical errors to facilitate machine reasoning.
• Vibe Coding in Critical Systems: While "vibe coding" may eventually apply to critical sectors, Traversal views its role as the essential safety net that validates and debugs AI-generated code when unexpected system interactions occur.
• Rapid Fire Insights:
  • AI Breakout Category: Applications leveraging reasoning models, specifically in diagnostics and healthcare.
  • Admired Figure: Demis Hassabis (DeepMind/Google).
  • Recommended Content: "The Bitter Lesson" by Rich Sutton.
  • Admired Agents: Glean, Perplexity.
  • Personal AI Tool: ChatGPT (used as a primary assistant).