◆ Supervised Worker
Pulls the populations, draws the samples, fetches the evidence, and runs the control test, going full-population where the data allows. Documents the workpaper with tickmarks and exceptions noted, logs every step and evidence link for re-performance, and routes the conclusions to a reviewing audit agent that gates them.
Memory
Working The control being tested + evidence gathered + exceptions so far.
Episodic Prior-period results for the same control.
Semantic Test procedures, sampling methodology, and what 'pass' means per control.
Procedural Evidence-pull recipes per system refined from reviewing-agent feedback.
Store Working-paper store + data extracts
Orchestration
orchestrator-worker MCP
Harness · Managed Agents … sandboxed code execution over data extracts; structured workpaper notes.
Tools
{ } Source systems (read-only extracts) API ›_ Full-population analytics Code exec ▣ Legacy GUI for evidence Computer use { } Working-paper system API ⇄ Reviewing audit agent A2A
Evals & guardrails
- Reproducibility: every test step + evidence link is logged for re-performance.
- Sampling methodology validated; deviations flagged, not silently adjusted.
- A reviewing audit agent must approve the conclusions … the agent tests, the reviewing agent opines.
Offline reflection
Learns which evidence sources are reliable vs. flaky per system and updates its pull recipes to fetch defensible evidence first time.
Frontier edge
- ▲Formal action-gating: pulls are read-only by construction and every evidence link is cryptographically signed and replayable, so a workpaper can be re-performed bit-for-bit by an examiner.
- ▲Self-improving fleet: an evidence-pull recipe learned against one legacy system propagates to every tester touching that system, so the whole audit population gets faster between runs.
- ▲Multimodal evidence capture: reads screenshots, scanned approvals and chart-based dashboards natively when a system has no API, so computer-use evidence is as structured as an extract.
A sample run
Trigger Fieldwork on access-recertification control for a trading system.
- 1Extract the full entitlement population and the recertification records.
- 2Test 100% for recertified-vs-active and segregation-of-duties conflicts.
- 3Document exceptions with evidence links in the workpaper.
Output A completed workpaper testing the full population, 14 exceptions flagged with evidence … gated by the reviewing audit agent for conclusion.
In numbers
100% of the book, continuously
Population coverage
21 min
Median control-test turnaround