System · Implemented
Agentic system
How Tessarion implements stateful orchestration, typed tools, memory, interruption, and controlled improvement.
What counts as an agentic workflow
Tessarion uses an agentic workflow only when a task needs state, conditional routing, tool selection, interruption, resume, or multi-step validation. Extraction, scoring, ranking, authorization, and persistence rules remain deterministic services.
Three workflow families are defined: Concept Intelligence, Learning Diagnosis, and Socratic Tutor. Each has typed state, allowed tools, bounded retries, terminal states, checkpoints, and trace events.
Tools and permissions
- Every tool has Zod input and output schemas.
- Workspace and user scope are resolved before execution.
- Read and write tools have separate permission and retry policies.
- Write tools require idempotency and produce an audit event.
- MCP exposure is derived from the internal registry; internal-only tools remain private.
Memory and checkpoints
- Working state is stored in workflow checkpoints.
- Teach-back and tutor sessions remain canonical in Postgres.
- Semantic memory is a derived retrieval projection.
- Relational memory is a derived graph projection.
- Hidden reasoning is not stored; structured decisions, evidence IDs, and safe summaries are stored.
Controlled improvement
Production workflows do not rewrite their own prompts or policies. A failure is sanitized, added to a reviewed dataset, reproduced offline, compared against a frozen baseline, and promoted only after human review.