GitHub pull request#1174
Add four-valued capability probe to provider adapters
- Sessions
- 18
- Outputs
- 7
Why Sabia
How agentic tools change the work.
Finance sees a bill. Engineering sees traces. Neither sees the work.
Sabia follows the Artifacts your team produces, as they stand in the suites where they live.
Grounded in Artifacts.
Your team does not produce one file per message. An Artifact is sessions and actions, accumulated. See how AI updates it over time.
Artifact
The document, pull request, issue, deck, sheet, or design that lasts.
Output
Each time that Artifact is created or changed.
Activity
The sessions, people, models, and tools behind those changes.
Those Artifacts live in the suites you already run: code and issues, Drive, Docs, Sheets, Slides, Word, Excel, PowerPoint, Figma, Notion, chat, and the rest of the work surface.
- GitHub
- GitLab
- Linear
- Jira
- Google Drive
- Google Docs
- Google Sheets
- Google Slides
- Microsoft Word
- Microsoft Excel
- PowerPoint
- Figma
- Notion
- Slack
Artifact Enrichment
The Artifact, as it actually is.
Sabia connects to your systems to see how agentic work evolves, so you see the full picture of the work.
01
Find the Artifact
Activity points at a pull request, document, sheet, deck, issue, design, or page.
02
Read the system of record
Drive, Docs, Sheets, Word, Excel, PowerPoint, Figma, Notion, GitHub, Linear, and the rest already hold the current facts.
03
Keep those facts current
Title, owners, reviews, and whether it shipped stay on the Artifact as the work moves.
04
Judge the work itself
Evaluations run against that Artifact, not the session that produced it.
See where AI consumption goes.
Tokens and dollars attach to model, person, source, and project.
- Model
- Person
- Source
- Project
Your Artifacts, judged.
Sessions and actions accumulate on one Artifact. Evaluations judge that Artifact as it stands.
Google DocOperating review
Q3 AI program performance, current draft
- Sessions
- 11
- Outputs
- 9
Excel workbookPipeline.xlsx
Sales forecast with agent-built scenarios
- Sessions
- 6
- Outputs
- 14
Word documentProposal.docx
Customer proposal, revised after legal review
- Sessions
- 8
- Outputs
- 5
PowerPointBoard deck
Quarterly operating review slides
- Sessions
- 5
- Outputs
- 4
Figma fileCheckout v4
Checkout flow, ready for engineering
- Sessions
- 9
- Outputs
- 6
Privacy by default
Collection follows the connection.
Usage connections capture consumption. Work-tool connections refresh Artifact facts so the work can be evaluated. What Sabia holds depends on what you enable.
Usage connections
- Token, request and usage quantities
- Cost records and model identifiers
- Timestamps and source metadata
- Project, workspace and member context when supplied
Work tools
- Artifact type, status and source reference
- Outputs recorded against the Artifact
- Current facts from the suite where the Artifact lives
- Evaluation results associated with the Artifact
Tenant isolation, enforced twice
Application permission checks, plus Postgres grants and row-level security. Every tenant row carries its organization.
Credentials never touch the exposed schema
Credential envelopes live in an unexposed schema, encrypted with AES-256-GCM and a per-record nonce.
Access matches the connection
Each provider uses the narrowest supported access pattern. Connections can be revoked, and ingestion keys are stored only as hashes.
How Sabia compares.
Trace and evaluation platforms follow each agent run. FinOps platforms allocate the bill. AI vendor dashboards cover their own products. Sabia follows the work itself: what AI produced, what it cost, and whether it held up.
Writing from Sabia on AI cost, output, and quality.
Compare Sabia with the alternatives