Gaia 3.1 release highlights

Gaia 3.1 spans releases 3.1.0-3.1.103 and focuses on a practical enterprise question: how do teams keep expanding agent capabilities without losing control of cost, evidence, identity, delivery, and the end-user experience?

This cycle adds an AI FinOps foundation, makes governance more executable at runtime, broadens document retrieval across backends and formats, strengthens versioned project and workflow operations, gives teams more control over public conversation channels, and hardens enterprise deployment and integration paths.

Major themes

• AI usage became a managed operating cost: cost budgets, workload policies, price evidence, usage attribution, billing reports, and model-routing visibility give teams a clearer path from token telemetry to financial control.
• Governance moved into the runtime path: external agent telemetry, review runs, trust tiers, framework mappings, and the MCP governance gateway connect policies and evidence to real tool execution.
• Document intelligence became more portable: PostgreSQL, Gaia-managed, and external retrieval backends can participate in a common architecture while Gaia retains folder semantics, access rules, structure, citations, and diagnostics.
• Projects and workflows became more release-aware: project versions, comparisons, branches, safer import and export previews, runtime service-account bindings, execution principals, and richer workflow traces make change easier to review.
• End-user channels became more configurable and safer: portal controls, custom fonts and avatars, feedback actions, localized messages, shared conversations, PII redaction, prompt-injection refusals, and browser-safe payload handling improve embedded experiences.
• Enterprise operations gained stronger foundations: Azure Marketplace activation and entitlement work, project-scoped credentials, Azure Key Vault hydration, secure image pipelines, model-catalog sync, operational status, and capacity evidence reduce deployment friction.

AI FinOps moved beyond token counting

Different models, voice services, tool calls, and realtime sessions do not have interchangeable costs. Gaia 3.1 begins treating that difference as an operating concern rather than a reporting detail.

The release adds cost-oriented budget policies, model price catalog contracts, workload policy approvals and expirations, usage and billing reports, CSV exports, and clearer attribution across organizations, projects, users, models, and service accounts. Runtime enforcement can persist a user-visible explanation when a configured boundary is crossed, while routing telemetry helps operators understand which model target handled a request and why.

This is an AI FinOps foundation, not a claim that cost management is finished. The important change is that cost evidence, policy, routing, and reporting now share a more explicit model.

Governance became executable

Governance records are useful only when they can influence or explain what an agent actually did. Gaia 3.1 adds stronger runtime connections between agent-system records, telemetry, evidence, reviews, and tool execution.

The MCP governance gateway can evaluate tool allow and deny rules, sensitive-tool approvals, rate budgets, schema fingerprints, drift, and response content around an MCP call. External agent systems can send OpenTelemetry evidence into the control plane, and governance reviews can use that evidence alongside linked policies, controls, and obligations.

The release also expands operational mappings for the EU AI Act and ISO/IEC 42001. These mappings help teams organize evidence and responsibilities; they do not imply certification or legal compliance by themselves.

Retrieval became cloud-neutral without giving up lineage

Gaia 3.1 introduces a clearer retrieval-provider architecture. PostgreSQL remains a compact default, Gaia-managed retrieval can serve deployments that want an internal engine, and external managed backends such as Azure AI Search can plug into the same product-level contract.

Gaia continues to own the parts that make retrieval operational: folder access, source records, structure search, query planning, ranking and reranking, citation lineage, diagnostics, and answer context packing. Multilingual query rewriting, accent-insensitive matching, richer backend diagnostics, and support for additional office and presentation formats extend that model.

The result is more deployment choice without forcing teams to surrender the evidence chain between a source, a retrieved passage, and an answer.

Delivery gained versions, identities, and better traces

Projects now have stronger versioning, comparison, and branch workflows. Import and export paths gained more complete previews, patch validation, and conflict handling so project change can be inspected before it is applied.

Workflow execution also gained clearer identity. Runtime service-account bindings and execution principals help teams distinguish who or what acted, while input and output mappings, prompt templates, lifecycle hooks, locking, and richer trace views make runs easier to understand.

Together, these changes move project configuration closer to a reviewable delivery artifact rather than mutable setup that is difficult to reconstruct later.

Public conversation experiences became product surfaces

Text channels and personal assistants received a broad set of controls for headers, avatars, bubbles, send actions, feedback icons, timestamps, composer layout, custom fonts, disclaimers, banners, close behavior, and localized messages.

The cycle also added shared portal conversations across tabs, embedded parameter handling, channel-scoped storage, PII redaction options, localized prompt-injection refusals, and browser-safe serialization for user-facing conversation payloads.

These details matter because an embedded assistant is often the part of the system customers see first. Teams can now shape that experience more deliberately while retaining the security and audit boundaries behind it.

Enterprise deployment became more repeatable

Gaia 3.1 strengthens the operational path around the product. Azure Marketplace activation and entitlement management, organization licensing, Azure Foundry model sync, Azure Key Vault hydration, project-scoped credentials, secure image deployment, signed release assets, maintenance mode, operational status, and repeatable capacity certification all received concrete implementation work.

The goal is not a single prescribed cloud topology. It is a more reviewable path from configuration and credentials through deployment, operation, and evidence.

The next 30 days

Over the next month, we will publish six deep dives into the major 3.1 themes:

1. AI FinOps and model control
2. Runtime governance and governed MCP execution
3. Cloud-neutral retrieval and evidence-ready documents
4. Project versions, workflow identity, and delivery control
5. Configurable and safer end-user channels
6. Enterprise deployment, integration, and operational readiness

What comes next

Gaia 3.1 gives enterprise teams more control as agent systems expand. Gaia 3.2 will focus on performance and productivity: reducing the time and friction required to operate the platform, complete everyday work, and move from intent to a verified outcome.