Intelligence Architect

What matters most

• Architect agent context

  The role decides what each agent needs to know and writes the procedures, skill documents and system prompts it consumes. Prepare to explain how you would divide accounting guidance across those layers and make instructions hard to misread.

• Set a cross-team quality bar

  Basis calls consistency across agents the role’s primary mission. It expects standards for procedure writing, useful skill documents and when context is ready to ship. Prepare a concrete way to review and version written context across product teams.

• Turn observed failures into revisions

  The role watches agents at work, identifies missing context when output falls short and owns the resulting writing changes; measurement is shared with product and evaluation teams. Prepare to distinguish a context gap from another possible cause before proposing an edit.

• Define good performance at each layer

  The posting asks the Intelligence Architect to define what good means and help design how it is measured, not merely judge the final answer. Prepare to describe what evidence would show whether a procedure, skill document or system prompt helped an agent complete accounting work correctly.

• Inform future product capability

  Insights from current agent behavior feed product decisions about new surfaces and higher capability for existing ones. Prepare to connect observed behavior to a bounded recommendation, while distinguishing demonstrated capability from a possibility as models improve.

Likely challenges to think through

• Precision without brittle instructions

  A likely challenge is writing guidance specific enough for reconciliations, tax, audit or analysis without making it fail when a case differs. For preparation, take one stated task and draft a short procedure that separates required steps, judgment calls and situations needing more information. This is a preparation exercise, not a stated hiring task.

• Finding causes in long-running work

  Because the agents perform hours-long, end-to-end work, a poor final result may not reveal which context layer contributed to it. Prepare a hypothetical failure trace: identify the first consequential deviation, compare it with the relevant instruction and specify what further evidence would test your diagnosis.

• Consistency versus task-specific needs

  Standards must hold across product teams, while different accounting tasks may need different knowledge and procedures. Prepare to explain which writing conventions should remain common and where task-specific exceptions would be justified; the posting does not prescribe that boundary.

Topics to prepare

• Agent context architecture

  The role decides what agents need to know and how procedures, domain knowledge, skills and system prompts fit together. For revision, sketch those layers for a hypothetical reconciliation agent and explain what belongs in each.

• Accounting-work context

  Basis names reconciliations, tax preparation, audit procedures and financial analysis as agent work. As preparation, choose one and map its inputs, decisions and expected output; use it to identify what an agent would need clarified by an accountant. This is a study exercise, not a specified accounting qualification.

• Performance and causal diagnosis

  The posting asks this role to define good performance at each layer and investigate why agent work succeeds or fails. For revision, take a hypothetical incorrect output and separate an observable error from possible gaps in procedure, domain knowledge or measurement.

• Context quality and consistency

  The role sets standards across agents, including when context is ready to ship; Basis says context is owned and versioned. As preparation, outline criteria for clarity, consistency and review of a proposed context change, without treating the criteria as Basis’s existing process.

• Agent behavior and product direction

  The posting says observations of current agents inform decisions about future capabilities. For revision, consider a hypothetical recurring limitation and explain what evidence would distinguish a context edit from a broader product opportunity.

Skills to revise

• Write an agent procedure

  Draft a short procedure for a hypothetical reconciliation, specifying inputs, steps, exceptions and completion criteria. Revise any sentence that could be followed in two different ways. Writing precise procedures is an explicit responsibility.

• Structure a skill document and system prompt

  Using the same hypothetical task, draft a compact skill document and system prompt, then remove duplicated or conflicting instructions. The role authors both and decides what each context layer says.

• Trace an output failure to a context edit

  Invent a flawed agent output, identify the specific instruction or missing knowledge that might explain it, and write a targeted revision plus a check for whether it helps. The posting assigns observation and context writing to this role, while measurement is shared with product and evaluation teams.

• Review context against a quality bar

  Create a short review checklist, apply it to your draft materials and record what you would change before release. This practices the stated responsibility to set consistent standards and decide when context is ready to ship.

• Clarify domain assumptions with collaborators

  Write five specific questions you would ask an accountant and an engineer about your hypothetical workflow, then mark which draft instructions depend on their answers. The posting explicitly calls for learning from both groups and making decisions amid ambiguity.

• Map the context layers

  The role owns procedures, skill documents and system prompts. Prepare a one-page diagram showing what each layer would contain, how they fit together and where ambiguity could cause an agent to misread instructions.

• Draft a sample accounting procedure

  As a preparation exercise, choose one task named in the posting, such as reconciliation. Write a short illustrative agent procedure with required inputs, decision points, expected output and cases that call for clarification; label assumptions rather than presenting them as Basis practices.

• Prepare a project walkthrough

  If you have relevant work to share, select one example of instructions or context you designed. Bring a before-and-after excerpt and explain your decisions, the observed behavior and what you can and cannot attribute to the edit.

• Practice diagnosing an agent failure

  Create a hypothetical output that falls short, then trace plausible causes across the procedure, domain knowledge and system prompt. Show the evidence you would seek before revising the text.

• Define quality and shipping criteria

  Draft a brief review rubric for clarity, consistency across agents and readiness to ship, plus a versioned change example. This is a proposed exercise, not a stated Basis checklist.

• Outline how to measure improvement

  For the sample task, propose observable success criteria at the instruction and end-to-end work levels. Note how you would collaborate with product and evaluation teams on measurement while owning the written changes, as the role describes.

• Bring questions about tradeoffs and collaboration

  Prepare questions on how accountants and engineers validate context, how teams handle conflicting instructions, and how evidence from current agents informs future product decisions.