Agent workspace
A workspace where AI agents do my everyday work, with shared memory and written rules.
It keeps my conversations and memory in one place, instead of scattered across different AI products. A method worked out once is written down and reused, so the workspace is meant to get more useful over time.
A coordinating session and several task-scoped specialist sessions share a file-based state. Each has its own scope, operating rules and memory.
What is in it
Gates. I write requirements and acceptance criteria before a capability is built. A later phase cannot start until I have tested the earlier one.
A searchable memory. A local knowledge base distils my own AI conversation transcripts, from several assistant products, into searchable summaries.
An audit that failed. I audited the knowledge base's automatic classification with a statistical acceptance-sampling plan. The batch failed the agreed threshold. I recorded the defect rate instead of calling it complete.
A model that rewrote decisions. A small local model turned rejected options into stated decisions when it summarised. I measured how often (28.6% on a hand-labelled set) and stopped it with deterministic checks in front of the model output.
Choosing by test. For local bilingual speech-to-text, I picked the model by testing two candidates on real recordings, and replaced the first choice when it failed on English and mixed-language speech.
Permissions enforced outside the model. A five-level permission model for automated actions. The high-risk levels are enforced by operating-system and configuration controls, not by instructions to the model.
Reviewing the reviewers. I asked two external AI systems for independent reviews of the same design questions, checked their claims against primary sources, and corrected errors in my own earlier decisions and in one reviewer's advice.