Fifty People in Slack. Four Minutes to the Smoking Gun.

Host Sonya Huang, joined by Sequoia partner Bogomil Balkansky, asks the hot-take question first: will DevOps or SRE as we know them even exist five years from now? Anish Agarwal, co-founder of Traversal, says the job will exist, but it will look fundamentally different. He reaches for healthcare. Stage one is a heart attack: nothing five minutes from now matters except solving it, which is a high-severity incident. Stage two is a chronic condition, a sprained ankle that makes it hard to plan three months out — the daily stream of alerts and the question of whether a deployment was safe. Stage three is life-hacking sleep and nutrients so you have a high-quality life, which is planning the next five years of infrastructure.

What people in DevOps, SRE, and on-call actually live like, he says, is having a heart attack twice a week and dealing with a debilitating chronic condition every day. Balkansky pictures the people he knows in scrubs, wearing CGM-style monitors to life-hack their infrastructure. If Traversal does its job, those first two stages — the high-severity incidents and the death-by-a-thousand-cuts — should be work that AI and AI agents take over. DevOps and SRE get the creative part: what infrastructure should look like for the next year and the next five. It becomes a much more fulfilling job.

Balkansky asks what Cursor, Windsurf, and autonomous coding mean for reliability. Agarwal's short-term answer is a tale of two worlds. Vibe coding is popular: a prompt or two and something stands up that people can play with. He compares it to fast fashion — try it on, like it or discard it, wear the next thing. In that world reliability barely matters. There is no craft. You create it and throw it away. Balkansky later names the poles: Louis Vuitton haute couture on one side, Shein on the other.

The other world is mission-critical systems in payments, financial institutions, security, infrastructure, and streaming. Enterprises Traversal works with are already using AI-powered engineering tools. The local unit tests pass. That piece of code looks perfect. Large-scale systems do not break there. They break when different pieces interact in a way nobody realized or could foresee. Because an AI wrote the code, humans no longer have the context. They did not write it, so they cannot debug it.

Unless the industry finds ways to use AI for software maintenance, Agarwal thinks it will get throttled. Either companies will disallow AI coding tools because there is too much downtime, or a new class of tools will grow up around maintenance. Traversal is betting on the latter. The 10x engineer who used to short-circuit a debug path because they wrote the error message themselves loses that structural advantage when the code is generated. You need an AI troubleshooter.

The lifecycle of an incident, Agarwal says, looks surprisingly the same across companies. A customer logs on, something is off, they complain to support. Support confirms it is not user error. The issue escalates in a game of telephone up five ladders of engineering sophistication until it hits DevOps and SRE, who decide whether the impact and immediacy make it worthy of an incident. Then chaos: a Slack channel, 30 to 50 people, implicit blame, a who-done-it. Almost without fail a 10x engineer who has been at the company a long time plays Sherlock Holmes, finds what happened by a path nobody can reconstruct, ships a hotfix or rollback, and a longer-term fix comes later.

What he finds incredible is the money around that ritual. Observability companies are typically the second-largest software spend after cloud, and root cause analysis is still in that state. The tools did the best they could with the technology available even six months ago. Observability is the creation of telemetry — MELT data: metrics, events, logs, and traces — plus storage and a visualization layer so people can put little eyeballs on slices of the system. That is where it ends, because that is all that could be done. The complex workflow of troubleshooting remains manual. An agent, Raj Agrawal says, is an LLM orchestration of tools: fetch logs, fetch metrics, format data, run anomaly detection. The hard part is defining those tools richly enough that an RCA can be expressed as a sequence of calls. The data is so large a single trace may not fit in context, so the agent has to feed the model step by step and slowly build a path to the root cause.

Early on they tried to mimic how an SRE debugs: look at one piece of evidence, decide the next. Humans make hops with system knowledge the agent does not have. Traversal leaned on scale instead. The Google SRE handbook's golden signals — latency, error rate — become a health-monitoring search over a rich space. Fewer inspired leaps, more sequential steps that are justified. The environment where that works is counterintuitive. They sometimes add less value at a Series A startup than at a large enterprise. Big companies have mature observability: the data is in place. Teams are fragmented, so no one person has enough context, which is why 30 or 50 people pile into the channel. The agent wins when the reasoning steps from trigger to root cause can be found in the data, just too much for any human to keep in their head. When the data is not there, the agent suffers. That tends not to be the case at the enterprise.

Balkansky asks for the self-driving analogy. Where the root cause is in the data — a log, a PR — Agarwal puts Traversal at L4, not L5. They can flag the problematic PR or the smoking-gun log, but the last mile of the fix is still open, especially when remediation is a system-wide change rather than a local file. That last mile is where coding agents get interesting. Where the root cause is not in the data, they are more like L2: important symptoms that help a human make the remaining hops. When they notice that gap, they tell the customer to instrument so the system is more observable to an AI.

Incumbents are not asleep. The heart of the opening is that observability is expensive, so it is extremely fragmented. Walk into a big enterprise and they are on Datadog, Splunk, Dynatrace, Elastic, Grafana, and ServiceNow — everything — in a game of attrition. Pricing is based on data stored, so company A is not incentivized to give you better insights from anything stored in company B. An on-call engineer calls on all five or six tools. That historical fragmentation, Agarwal says, is what gives a company like Traversal a chance: for now, they can stay agnostic to where the data lives.

They started the usual way: small companies, the last 100 Slack incidents, a human-summarized meta-workflow stuffed into the then-popular ReAct framework. It worked. Then they hit a large company with thousands of microservices. They ran the system on a historical incident. Accuracy was 0 percent. Prompts did not move it. Nothing moved it. It was stubbornly at zero. Agarwal remembers the week clearly. He and Agrawal had a Negroni. When they came back they made two decisions: nothing about a specific company would be hardcoded into the prompts, and nothing about a human workflow would be hardcoded into the agent. The complexity had to go somewhere. It went into computation — spending tokens at inference time. Once they found an architecture that exploited that, accuracy started shooting up. If the answer lies in the data, they now get there more than 90 percent of the time, in two to four minutes. Humans in the Slack channel spend most of their time verifying. Month-to-month time-to-resolution dropped, as did the average number of people in the channel — the two numbers any enterprise cares about.

The gold standard for accuracy is live incidents. Customers often have two or three a week; after a couple of hours you read the postmortem and evaluate. They also score higher-volume subtasks, such as searching observability data for a specific fact. Under the covers there are two phases. Offline, they learn a dependency map: LLMs traverse the codebase and logs to understand how tags and functions relate; statistics pull causal traces out of natural variation in time series, the same family of methods the founders worked on in graduate school; self-play prioritizes promising paths. Online, when an incident arrives, the agent uses real-time signals plus that map to choose hops. They only ask for read-only access, because enterprises do not want another tool generating more data. The offline pass takes about five to ten hours, a day for larger customers. A single investigation can make thousands of network calls.

Architectural constraint number two is the enterprise LLM contract. If you bring your own model, you sit in security review for about a year. Traversal ties its hands: use whichever model the customer already has, typically OpenAI or Anthropic. Most of the complexity is the tool set the LLM orchestrates. Fine-tuning, when it happens, happens inside the customer's environment — point at their Azure OpenAI instance, compare how far the agent got from the pinned root cause after each incident, and shrink the gap over time.

Agrawal's PhD work with the Broad Institute studied CRISPR interventions on gene regulatory networks: knock something out, watch how expression percolates. Swap genes for microservices and a PR change for a knockout, and the causal-structure problem is almost identical. He calls it sheer luck. They did not know it would transfer until they were in the weeds. Other surprises, Agarwal says, start with where innovation lives. Coming from academia, on a path toward becoming a professor, he assumed the frontier was the university. What he found is the industrial age of AI: the most interesting work is happening at small, research-focused startups. Enterprises are hungrier than expected. Deals that would have taken a year or two close in a couple of months.

Future-proofing is a constant challenge. If you are in product, design, or core engineering, you have to bet where AI will be six months from now and be willing to re-evaluate everything. The good news is it only gets better; the product is not going to get worse. One bet that paid off: last September they saw reasoning models coming and architected the system so those models would shine. Companies that string together a few of those six-month bets will win. Asked whether he would rather be an AI team or an observability team taking this problem, he picks the AI team, because you get more fluent at making those bets.

The company is about 90 percent engineers. A few have machine-learning PhDs; a large fraction come from traditional software engineering and got hooked on generative AI. The barrier to building agents is lower than the old days of writing gradient updates. They hire infrastructure engineers specifically to let the agent do more work than a human can — those thousands of network calls per investigation. Product managers were about to join. What unifies the group, Agarwal says, is excitement for generative AI and a willingness to throw away what did not work. Titles are useless. The test is how experimental someone is: little bets, hypotheses, put it to the test. Quick iteration with AI and data is a mindset, not a credential. Balkansky notes that none of the four co-founders came from observability, and the company is heavy on AI and light on domain insiders — and they are still finding smoking-gun logs that used to take little armies.

What does a future observability or SRE team look like? Agarwal's first answer is that AI systems themselves will break in creative ways. Good SRE principles plus fluency in how LLMs fail is what will make someone valuable. The data layer will change too. Part of being good at the job is knowing how to write a log: not flooding the system, just the right amount. Logs will look different because they are no longer meant to be scrolled by a human. They are meant to be consumed by an AI. Balkansky asks how Cursor rates at logging. Agrawal thinks it is already better than a human — it formats cleanly, and humans make typos — but it is still mimicking traditional logs. You want to bury as much information as possible in the message field, the meat the model uses to choose hops. People resist that because a human cannot read a long stack trace. We have shorter context windows than LLMs. The remaining art is connecting logs to business logic. "Fine" for an engineering system is a business definition. Cursor cannot do that unless it fully understands the business.

If everything Traversal is building goes right, will engineers at the largest banks, payments companies, and healthcare institutions vibe-code their programs? To some extent, yes. Software engineering will be judged on whether the code fulfilled a function, typically a test, not on how it was written. That is efficient. It is also the failure mode he already described: systems fail in unexpected interactions, a third party misses an SLA, and you do not know what was in the file. Humans shine today because they wrote the code and can short-circuit. Generated code removes that advantage. You really need an AI troubleshooter. That is Traversal's reason to exist.

Rapid fire, mostly one sentence. Breakout AI application category in the next 12 months: anywhere reasoning models get used better, diagnostics especially, healthcare as the prime example. A founder they admire: Demis Hassabis, for foundational work and a habit of taking the hardest problems. A five-minute way to get smart on AI: Rich Sutton's essay "The Bitter Lesson." Agent companies they admire: Perplexity, and Glean among the obvious names. Personal AI they actually use: both founders say they use nothing but ChatGPT; Agrawal calls it his best friend. Will we be vibe-coding banking, payments, and healthcare in five years? To some extent, yes.