Agent engineering starts with the business environment

Two business processes built as agents. A support agent resolves a failed replacement and can refund only when policy, a manager and the customer allow it. A review agent turns conflicting contract documents into a cited exception memo for the people who decide.

Recorded demonstrations on public and authored records with simulated participants; not client work.

Business agent demonstrations

  • Investigating a failed replacement

    A replacement order never reached the customer and the return window has closed. A support agent investigates the order, establishes whether a refund is allowed and resolves the case with the customer.

  • Reviewing a supplier NDA before a pilot

    A procurement analyst needs a supplier NDA reviewed before an analytics pilot. A review agent prepares an internal exception memo that cites its evidence and names the decisions people must take.

How we build and evaluate a business agent

  1. Specify the business result

    Identify the request, current process, responsible people, source systems, permitted actions and acceptable handoffs. Define what a useful completed result must contain before choosing a model or framework.

  2. Gather documented and tacit knowledge

    Collect policy and operational records. Ask process owners how they handle missing information, exceptions and conflicting instructions. Record decisions and unresolved questions explicitly; the demonstrations use labelled authored stakeholder exercises.

  3. Select and adapt the environment

    Check what an existing resource actually provides: documents, stateful tools, other actors, task definitions or graders. Inspect source rights and execute reset and evaluator controls before depending on it. Add the business semantics that are missing.

  4. Build the model-controlled tool loop

    The model decides what to investigate, receives real tool results and revises its next action. Code enforces identity, permissions and bounded resource use. Requests requiring a person remain visible handoffs.

  5. Judge outcomes and failure behavior

    Check the resulting business state and evidence separately from the agent’s narrative. Exercise wrong targets, policy changes, missing authority and lost responses. Comparative performance requires its own reviewed protocol.

  6. Present inspectable evidence

    Present the request and completed result first, then the investigation, independent checks and limitations. Recorded variations show how a changed customer request, lost response or revised policy affects the result. Visitors can inspect these retained executions without starting an agent or tool.

AI Workflows & Agents

From discovery to operation

  1. Discovery

    • Agreed required results and examples
    • Source inventory and gaps in documented or tacit knowledge
    • Decision owners, action permissions and unresolved questions
    Shown in these cases
    Authored discovery exercises translated into the two process/permission designs
    Not shown
    Actual client interviews or agreement on a client process

  2. Bounded pilot

    • Adapters within agreed access boundaries
    • Resettable evaluation environment and frozen acceptance cases
    • Working agent, independently inspected results and observed operating costs
    Shown in these cases
    Pinned native Support reads, custom synthetic ledger and recovery; Source-bound contract tools, actual memos and business-intent follow-up
    Not shown
    Production system adapters, representative completion rates or a measured economic advantage

  3. Operational rollout

    • Acceptance decision and staged access
    • Monitoring, exception ownership and recovery procedures
    • Review/correction measurements and maintenance responsibilities
    Shown in these cases
    Local state restrictions, durable receipt/reconciliation and source binding
    Not shown
    Client rollout, production service levels or deployed monitoring
What the environment survey changes

Illustrative environments and retained executions, not client case studies or public agent services. The source survey covers 81 purposefully selected resources; its catalog, source inspections, component trials and full runtime executions are distinct evidence. Source and computational research remain later, separate capabilities.

Read the full environment survey, source audit and selection method

  • tau retail components

    Used selected native retail reads and records, with independently executed tool/state/reset and DB-grader controls. A foundation for the Support investigation world.

    A return request is not a financial refund. Replacement history, company policy, authority, consent and durable refund storage are custom adaptations; no full benchmark reproduction.

  • Hyper-tau

    Informs requirements discovery: gather business evidence and stakeholder knowledge before constructing and testing an agent.

    Source-inspected method reference. The multi-role construction campaign was not executed.

  • ContractNLI

    Separates NDA statement interpretation from evidence spans. Informs component evaluation for Contract Exception Review.

    Source/schema and licence inspection only. A fixed hypothesis classifier does not perform procurement review, establish amendment authority or evaluate company policy.

  • AppWorld

    A candidate when work genuinely requires stateful APIs across several applications.

    Not adopted or run here. Protected app/task material has additional redistribution conditions; process semantics and outcome checks still need mapping.

  • The Agent Company

    A reference for multi-service workplace tasks and interactions with simulated colleagues.

    Not deployed here. Its broader workplace does not remove the need to specify company rules, permissions, resets and task-specific checks.

  • AgentDojo

    Informs tests where hostile instructions arrive inside lower-trust documents or tool results.

    Security evaluation is one component of a useful workflow. The external attack environment has not been reproduced here.

Earlier benefits evidence reference: Coverage Auditor — SBC Evidence Audit