Introduction
Hello, Jose here. This time, I joined Anthropic on the second day of its two-day event focused on developers and founders using Claude Code.
The conference was organized into a fixed schedule following three tracks:
- Founder stage (founders using Claude to solve industry problems)
- Builders stage (people with little or no programming background shipping software using Claude)
- Workshops (Hands-on sessions about the Anthropic ecosystem).
There are other excellent reports of the event ([1][2][3]), so this time I’ll write a collection of lessons from the talks I joined.
Most AI projects by large corporations end in failure
According to a 2025 MIT study, 95% of AI projects by major corporations yield zero return. RAND reports an 80% failure rate, while Gartner reports that more than 60% of AI projects were abandoned by 2026.
The presenters at Tsukumo Labs’ talk “The last mile is the spec” argue that part of the problem lies in the lack of specifications. And as a tool to solve that problem, they are developing Tsukumo Log, an app that can use computer logs to reproduce and automate tasks (sending emails, writing and running scripts, etc).
Some of their findings during the implementation were:
- With modern AI (LLMs), you can usually use computer operation logs to reproduce and automate operations, but high quality test data is critical: what are you feeding Claude and how can you interpret the results it gives?
- In agentic development, we should focus on the inverse process: define the expected answers given by Claude and backtrack them with a myriad of hypothesis..
Should Claude manage everything?
In the old days of software engineering (five years ago), the business side waited for code to be written (Can engineering get to this by Friday?). Yet now, code can be written at neck-breaking speed. The problem has shifted to decision-making: what do we ship and who should be responsible for that shipment?
Myrealtrip’s CTO, Wonjin Hur, experienced this first-hand when her team shifted to LLMs.They were working on a voice agent to handle customer calls asking for airline refund policies. Originally, they left Claude handle the whole workflow: parsing the data, computing the refund, and answering customers in natural language. Yet, this opened their system to a new class of hard-to-debug bugs.
So they changed their focus. They separated the deterministic Python code for parsing and calculating the refund amount and let Claude handle the calls using that information.
The results are revealing: they reduced expensive (and unnecessary) reasoning, with the added benefit of predictability for their business logic.
Sometimes Claude isn’t wrong; the client is.
In their book Pragmatic Programmer, Andrew Hunt and David Thomas argue we should actively use assertions to check for the impossible. This is exactly what happened to Gahee Seo at Federation. She described an unexpected challenge of building an HS Code classifier: customers could write in terms that only they understand! Misclassifications occurred because the input was ambiguous. The lesson? Build an input quality gate or put a human in the loop for ambiguous input.
Final thoughts
The event was great. We had a chance to talk to Anthropic engineers coming from San Francisco, and the banquet was delicious. Anthropic had a good sense for the omiyages, too!
