North American hedge fund · $1bn–$5bn AUM · 25–50 staff
A practical Claude Code program used the firm's own research workflows to teach scoping, source verification, testing, multi-agent design and the operating decisions behind shared agents.
The learning path
Define. Build. Verify. Operate.
The aim was practical fluency: understand what the agent should do, where it can fail and what changes when several people need to rely on it.
The investment team already understood the basics of using AI. Its next questions were more demanding: how should an analyst move from chat into a repeatable research agent, what data should it reach and how should the team test an answer that sounds convincing?
Altitude designed the program around the fund's own research materials and a real diligence ambition. Participants saw a bounded workflow built from their brief, then examined increasingly sophisticated systems to understand what made them reliable, inspectable and suitable for wider use.
Staff did not need to become software engineers. They needed to define an agent clearly, challenge its work and recognize when an experiment had become an infrastructure or deployment project.
From idea to operating system
The training moved from a workflow the analysts already understood to the controls required when that workflow is shared, scheduled or connected to more sources.
Agent-building progression
Define · build · challenge · operate
Choose a recurring research task the team already knows how to judge.
Define the result, permitted evidence, review point and visible failure modes.
Keep the first workflow narrow enough to inspect, test and improve.
Check sources, calculations, conflicts, missing data and regression tests.
Address sharing, versioning, permissions, monitoring and deployment.
Facts + checks
Evidence pack
Sourced figures, periods, units, calculations and visible gaps
Relevance + evidence
Signal filter
Relevant developments tied to research questions with source links
Multiple agents
Diligence system
Specialist agents, verification and one orchestrated review point
The objective is practical fluency: know what the agent should do, what evidence it used and where a person must decide.
Human judgment retainedStart with the real ambition
The firm's broader idea joined public filings, licensed research, internal notes, market information and event data with dashboards and scheduled agents. That was a useful destination, but too broad for a training exercise.
Altitude narrowed the first task to new-event diligence. When something happened to a covered company, the workflow would gather the relevant evidence, compare it with the existing research view and show what still required analyst judgment.
The distinction mattered. A bounded agent can be built around accessible sources and a familiar recurring task. A sector-wide system requires a data foundation, entitlements, deployment design and continuing operation. The training gave the team a practical way to recognize that boundary.
Session one
The first session introduced Claude Code as a working environment for repeatable agents. The team used Altitude's proprietary agent-building method without turning the session into a lesson about prompts or software tools.
In practical terms, the method required the builder to settle the desired result, permitted evidence, role of the firm's judgment, testing approach and delivery point before development began. The client's diligence workflow made those choices concrete.
Participants left with a simple take-home file that guides Claude Code through the scoping discussion and produces a written brief. It was deliberately easy to use in a normal working folder, without specialist setup.
The training adapted
The live portion of session one tried to run several builds at once. Participants spent too much time waiting for workflows to finish and had less opportunity to understand the design choices or challenge the results.
The client said the second half had lost some people as the session moved between running workflows. Altitude changed the format. For session two, the agents were built and running before the training began.
The room could then focus on the questions that determine whether a firm can rely on an agent: Is the output right? What happens when the agent cannot tell? What does each run cost? How does it operate without its original builder? How do several people collaborate without changing it unpredictably?
Session two
Three prepared systems increased in complexity while staying grounded in the client's research domain.
Company filings and permitted market information became a compact evidence pack. Figures retained their period, units and source; calculations showed their arithmetic; conflicts and missing data stayed visible.
Deterministic checks and hand-verified regression examples made silent extraction changes easier to catch.
A defined source set was screened for developments related to specific research questions and tripwires. The system aimed to reduce noise while keeping a direct path to the original source.
Relevance was judged against the research view rather than popularity or generic sentiment.
Specialist components examined filings, market information, management commentary, company materials, events, licensed research and the internal view. A verifier checked the evidence before an orchestration layer assembled the review.
Each specialist had a narrow responsibility, separate test surface and explicit human decision point.
These systems were training demonstrations. They were used to teach design and evaluation; they were not production deployments at the fund.
Operating fluency
Legal participated, so data access, responsibility and permitted use were part of the build discussion.
How should two analysts collaborate on the same agent without overwriting each other?
When is a local workflow appropriate, and when is a shared cloud environment required?
How do we detect a scheduled output drifting from its original format or standard?
How narrow should a specialist agent be before the task needs several components?
Which sources and actions require tighter permissions, legal review or monitoring?
How does the workflow run when the analyst who built it is unavailable?
What the team gained
Participants used a structured method to turn an investment workflow into a written agent brief, then examined sourced facts, conflicts, missing information and deterministic checks.
They worked through specialist and multi-agent designs and discussed cost, sharing, deployment, monitoring, permissions and collaboration.
The sessions also clarified the line between what staff can prototype themselves and what requires a properly scoped build. One participant requested access to a demonstrated workflow, showing that the examples connected with the team's work.
The team received reusable materials and recordings to continue applying the methods to its own research work.
Evidence boundary: the engagement established practical agent-building fluency and reusable training assets. No sustained-adoption, time-saving, production-deployment or investment-result claim is made.
The service behind this work
Role-specific training, live-work clinics and power-user tracks that build practical capability across the firm.
See how AI training works →This engagement also touched Managed AI Agents.
Read next
AI training
Three working sessions on specialist extraction, Excel model updates and agents the investment team could test.
AI training
Investment, technology, operations and compliance worked through live workflows, reusable skills and the decisions behind a coordinated programme.
Start with a recurring task the team understands, source material it is permitted to use and an output experienced staff can evaluate. Build the training around the work your firm actually wants to improve.