Self-improving workflows
A self-improving workflow is one the agent keeps editing while it is in use. The workflow itself is a set of instructions, skills and scripts. After each run the results are checked, and a failure that shows up more than once is treated as a fault in the workflow. The agent then edits the workflow so that failure stops happening, by adding a skill, adding a small script or correcting an instruction. Every edit is tested on the same fixed set of past cases before it is used, so its effect on the failure rate is measured each time.
The same 300 cases, every passNothing came back wrong
- 19%Pass 1AnalyseThe review agent adds a skill
- 5%Pass 2RerunAnalyseThe review agent adds a small script
- 1%Pass 3RerunAnalyseThe review agent changes the instructions
- 0%Pass 4Rerun
workflow.mdSeven steps, three by the review
Reconciliation
Cash and positions from three custodians, checked against the firm's book.
Due diligence questionnaires
180 investor questions, with every figure traceable to a source.
RFPs and tenders
60 tender sections, each scored against the buyer's rubric.
Why it compounds
One fix closes a class of errorA person who fixes a break fixes that break. A review agent reading 100 of them can work out which class they belong to and close the class. The charts above fall in steps for that reason.
Every change is measured before it shipsA proposed change is replayed against the frozen benchmark first. Rejected changes are recorded with the score that rejected them. A cheaper model that scores 6 points lower does not ship.
The person handles less over timeThe checkpoints stay. In the first week a person sees everything the process could not finish. By the last they see the decisions that needed a person anyway.