Anthropic Engineer With 20 Years of Experience Builds Observability for Claude Agents

Anthropic has an engineer on its team who has spent twenty years on a single question: can you trust the numbers? Gurmit Saran, an IEEE Senior Member, has worked on Anthropic's staff in San Francisco since 2025, building infrastructure to understand how Claude behaves. As he puts it, the goal is that data must be accurate, structured and trustworthy, and private information must stay private. IEEE Spectrum published the profile on 12 October 2026.

From a single metric definition to auditing agents

Saran's career began in 2007 at Yahoo's Bengaluru office, where he was tech lead of a data team. There he ran into a problem that kept recurring at every company. Different teams calculated "daily active users of the site" using slightly different logic and got different numbers. Leadership made decisions about resources and products without knowing which number to believe.

The team built a central repository of rigidly defined metrics. Saran specified which events count toward the calculation and how to handle edge cases where standard code breaks. On top of that came Yahoo! Digits, an internal analytics tool used by all of the company's products.

At Intuit, which he joined in 2010, Saran built analytics for TurboTax and QuickBooks. The company also launched a feature there for detecting fraudulent tax returns in real time. The system had to protect taxpayer data and meet IRS requirements. From that project he took a principle that still shapes his work: privacy cannot be bolted onto a finished system; it has to be built into the architecture from day one.

Saran arrived at Meta, then still Facebook, in 2016, when the company admitted it had overstated video view counts. He sees the cause in the architecture of the metrics. His answer was Metric360, a system in which a single definition governs all view calculations. By 2021, when he left for Wealthfront, it was in use across the company.

Why agents are harder to observe

At Anthropic, Saran applies the same approach to a different object. He compares the work to a doctor making decisions based on instrument readings: if one instrument drifts slightly, the error shows up in borderline cases, where the decision is hardest. AI is the same, he says: you have to make sure the instruments show the truth and that the data means what it claims to mean.

Agents make the task harder. The system executes sequences of actions, and the engineer has to record three things:

• what exactly the agent did;
• what information it based its actions on;
• whether the result matched the user's expectations.

All of this has to happen without exposing other people's data. According to Saran, the systems are designed to recognize sensitive information and protect it during processing. He calls this a far harder observability problem than any he has faced before. The industry has decent tools for databases and web services, but for LLMs, he says, they are only now being invented.

The role of standards and what to take from this story

Saran joined IEEE in March 2025 and sees the organization as a potential author of standards for AI infrastructure. His argument is simple: standards exist for building codes and software testing, but there are no equivalents yet for AI auditability. His second argument concerns access. According to him, IEEE operates in countries where engineers have no easy path into the global professional community, and that matters as AI development moves beyond a few organizations in a few cities.

The article does not disclose the architecture of Anthropic's internal systems; it is a profile, not a technical report. But its practical takeaways apply to any team running agents in production. First, fix single definitions for quality metrics so that different departments do not count a "successful response" differently. Then log the agent's chains of actions, not just its final answers. Sensitive-data masking should be built into the telemetry pipeline from the start, not added after the first incident. Saran considers auditability one of the main engineering challenges of the coming decade, and in his view it gets noticeably less attention than model capabilities.