The Agent System Running La Bandera Wasn't Built Around One AI Model

Context

Asking a client to adopt an AI-driven approach to how their engineering gets done is a hard sell on faith. The credible way to make that case was to run the firm's own operations on it first, for real, before ever proposing the same approach for someone else's team.

Decision

Deploy the system against the firm's own repositories and its own daily operational decisions first — real stakes, not a sandbox — and let it hold up under real use before proposing the same approach to a client.

Outcome

In the months since, the same strategic bet — building around multiple models instead of locking into one — has started showing up at scale industry-wide. Spotify's engineering team published a public account of a comparable agent working repo-wide in production. Shopify went further in public detail: their own account describes an internal platform built specifically so multiple frontier models can operate side by side rather than standardizing on one, with the resulting agent now credited as a co-author on roughly one in eight of everything the company merges.

Read plainly: this isn't an isolated bet. Multiple top engineering organizations are independently arriving at the same conclusion about avoiding single-model lock-in, on their own timelines and for their own reasons.

The Judgment Call

The common approach teams are taking right now with AI-assisted engineering: get developer licenses for one specific frontier model and build workflows around it — fast to start, but locked into whatever that model happens to be strong and weak at, with no easy path to adapt when a better option shows up next month.

It's tempting to assume a protocol layer solves this on its own — stand up a standard integration server so one vendor's agent can call a tool, or even another vendor's agent, and call the lock-in problem solved. That gets a model access to more tools and other agents. It does not get consistent behavior, standards, or judgment across those models, which is a materially different and harder problem than wiring endpoints together. And that's before counting what every one of those tool definitions costs in tokens on every single call — overhead that compounds with each server bolted on, not something a simple integration count captures.

The alternative, harder to build up front: an agentic system designed from day one to draw on more than one model rather than commit to a single vendor's roadmap, capturing the best of what each does well rather than inheriting the tradeoffs of just one. Chose the harder path — the whole point was not getting stuck with one tool's blind spots six months in.