Wealthsimple CTO Diederik van Liere says teams that cleared out old instructions for AI models saw token use fall by as much as 50% on some tasks, while performance improved. His broader message to developers and businesses: when models change, yesterday’s helpful scaffolding can become today’s expensive clutter.
Watch Desk analysis
What happened
In an interview with Digital Journal, van Liere describes asking Wealthsimple’s engineering teams to review the README and instruction files they had built to guide models. The teams tested the same task with and without the accumulated guidance. Van Liere says token use fell “sometimes by 50%” and performance went up. He now tells teams to start from scratch after a major model update, then test what the newer model actually needs. Read the Digital Journal interview.
The interview also describes how Wealthsimple is putting AI into internal workflows. Its Magic platform lets employees build small tools connected to company data, while a Slack agent called Compass answers business questions from the data warehouse. Van Liere says the company has thousands of these Magic sites, including tools used by its business operations team.
For agents that query the warehouse, Wealthsimple currently uses read-only access. The company is developing Orbital for building and reviewing agents, and CoWatch, an internal beta tool that tracks agent tool calls. Van Liere says tracking responsibility for what an agent does remains an unsolved problem.
Why it matters
There are two useful lessons here: model updates can make old instructions wasteful, and easier access to AI makes reliable data and carefully limited permissions more important. The first affects operating costs as well as output; the second matters when agents are allowed to query company information or act on it.
These are van Liere’s account of Wealthsimple’s own experiments and systems, not a universal guarantee that deleting instructions will improve every model or workflow. But the advice is practical: test against the current model rather than assuming last quarter’s prompt handbook is still helping.
Our read
“Burn down the scaffolding” is a bracingly memorable way to describe routine maintenance. Teams should try it when models change, but measure the result on their own tasks rather than treating a reported saving of up to 50% as a forecast. And if an agent gets access to company data, read-only permissions and a record of its actions are sensible places to start. The agent era still needs some rather unglamorous filing systems.
What to watch
- Whether Wealthsimple’s reported token and performance gains hold across more tasks and model updates.
- Whether the company completes its data-warehouse migration and expands agent access.
- How Orbital’s review process and CoWatch’s activity records develop, and who is accountable for an agent’s actions.
Discussion spark: When a model changes, should teams start with a clean slate or keep the instructions they already know?
Sources and evidence
- Wealthsimple's CTO has his team burn down its AI scaffolding – Digital Journal (19 November 2026, 20:11 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.