Tessera: An Open-Source AI Agent Layer That Refuses Answers Without Proof

Tessera is an open-source project tackling a core weakness in enterprise AI agents: not model quality, but the inability to verify answers. It ingests heterogeneous sources — database tables, documents, logs, tickets — resolves entities that appear differently across sources (like 'Acme Corp' vs. 'acme-prod') into one unified knowledge graph, and traces every sentence of an answer back to the underlying records. Claims that can't be proven aren't fabricated; they're refused outright.

The same discipline extends to actions: an agent drafts a proposed action where every field ties to a verified claim, shows an exact dry-run preview, and only executes after human approval — leaving behind a receipt. Over MCP, Tessera enforces this trust contract across four layers: read-only evidence tools, propose-and-approve drafts, payload previews, and gated execution. Notably, no model vendor sits inside the trust path — an optional LLM may narrate results, but never attests to their truth.

One engine powers two reference use cases: a Business Data Copilot for querying structured enterprise records, and a DevEx Copilot for diagnosing failed CI/CD pipelines from logs and diffs. The project also ships a benchmark harness that turns trust into a measurable 'faithfulness' score tracked in CI, and is built with SAP AI Core and HANA Cloud compatibility in mind, signaling a clear on-prem, enterprise deployment target.