Agent Launcher
Turning repetitive analysis tasks into reusable agents, without being locked into one model or provider.
A platform to compare and run agents across different models, with executions persisted and auditable. Built for an investment management firm.
The problem
At an investment management firm, much of the research work is repetitive: the same questions about different companies, the same cross-checks, the same kind of summary.
Automating that with AI is tempting, but it has a catch: each task ends up tied to one specific prompt, model, and provider. When the provider changes pricing, models, or terms, the work has to be redone.
What we built
An adaptation layer between the agents and the various model providers. The user picks agent, model, and provider separately, and can change any one of the three without redoing the other two.
Every execution is persisted and audited, so it is possible to review what was asked, with which model, and what came back. In a setting where decisions carry financial consequences, being able to reconstruct how a conclusion was reached matters as much as the conclusion.
The agents are also exposed over an API, so other systems can consume them without going through the interface.
How it evolved
In later phases it stopped being only an agent launcher. It gained conversation with context, visualization, specialized agents, and the principle of using the organization's internal data before falling back on web search.
The MVP was technically validated and became one of the foundations of a broader AI architecture inside the client.