AI Product Updates, July 29: Gemini Goes Proactive, MCP Goes Stateless, and Agent Tools Multiply (2026)

At a glance

Google made two moves toward ambient assistance on July 29, 2026 (GMT+8): Gemini Spark began expanding to Google AI Pro subscribers in India as a background agent for Workspace tasks, while the Gemini macOS app added voice dictation, screen-aware reasoning, rewriting, and image editing. Google Pay also launched an opt-in Gemini finance assistant in India. For developers, MCP 2026-07-28 is the major infrastructure change: the protocol is now stateless, request-routable, and backed by a formal deprecation policy. GitHub, Immersive Labs, and several launch-day developer tools filled in the security and agent-operations layer.

Product or platform	Releasing entity	Release date	Core capability or change	Availability
Gemini Spark	Google	July 29, 2026	Background agent for Gmail, Docs, Sheets, calendars, and other directed tasks	Expanding over the next few weeks to Google AI Pro subscribers in India; Google says it is available to AI Ultra and Pro subscribers
Gemini for macOS	Google	July 29, 2026	Fn-key voice dictation, screen-aware reasoning, rewriting, and image generation/editing	Rolling out globally in English
Ask Google Pay	Google Pay	July 29, 2026	Gemini-powered spending analysis, savings guidance, and financial education	Opt-in launch in India; supports 10 Indian languages
MCP 2026-07-28	Model Context Protocol community; AWS AgentCore support	July 29, 2026 in the GMT+8 window	Stateless requests, header-based routing, cache hints, MRTR, auth hardening, and formal deprecations	Spec and Tier 1 SDKs released; AgentCore Gateway support is opt-in
CodeQL 2.26.1	GitHub	July 29, 2026	Better framework coverage and fewer Rust false positives	Automatically deployed to GitHub code scanning; GHES support is planned for a future release
Agentic Harness	Immersive Labs	July 29, 2026	Tests autonomous AI safety and token spend	Launch announced; product page directs teams to request a demo
Prelint	Prelint	July 29, 2026	Reviews pull requests against ADRs, documentation, and past decisions	Launch-day Product Hunt listing; the page shows free options
MemoryCustodian	waittim / MemoryCustodian	July 29, 2026	Local, repo-native memory for Codex, Claude Code, Gemini, and other coding agents	Launch-day Product Hunt listing; open source and free
/mission for Claude Code	Spine / Medley	July 29, 2026	Turns an outcome into a live graph and coordinates Claude Code and Codex workers	Launching today as a free Claude Code plugin; BYOK through OpenRouter
Framer AI Agents	Framer	July 29, 2026	Agents on the canvas, external AI connections, branching, and team publishing	Launch-day Product Hunt listing; available through Framer's product surface

Google moves from chat to background work

Gemini Spark

Gemini Spark is Google's clearest July 29 shift from a reactive assistant to a background agent. The service can work while a laptop is closed or a phone is locked, connect to Gmail, Docs, and Sheets without setup, and follow recurring instructions such as organizing travel confirmations into a spreadsheet or finding events for an open calendar slot. Google says Spark is driven by Gemini 3.6 Flash and asks for confirmation before high-stakes actions such as sending email or spending money. 1

The India rollout is staged. Google says Gemini Spark is expanding to Google AI Pro subscribers in India over the next few weeks, while the same announcement describes the service as available to Google AI Ultra and Pro subscribers. That is a meaningful distinction from a general free-tier launch: the capability is tied to paid Google AI plans and is not presented as a fully open consumer release. 1

For teams testing agentic office work, the useful questions are operational rather than rhetorical. Which connected apps can Spark read? Which actions require confirmation? How are recurring instructions stopped or audited? The launch post gives examples and guardrails, but not a public API or an enterprise administration surface.

Gemini for macOS

The Gemini app for macOS now responds to a long press of the Fn key. In its default mode, it turns speech into cleaned-up text at the cursor, removes filler words, and handles mid-sentence corrections. With Gemini reasoning enabled in settings, it can use selected files, images, documents, or on-screen text to summarize, rewrite, generate images, and edit an existing illustration. 2

Google says the voice capability is rolling out globally to all Gemini macOS users in English, with more languages to follow. The important product boundary is that dictation is broadly available, while screen-aware reasoning is an opt-in setting. Developers who work across terminals, editors, and design tools can try the workflow without switching windows, but the feature is still a desktop-app interaction rather than a public automation API. 2

Ask Google Pay

Google Pay launched Ask Google Pay in India as an opt-in conversational layer inside the payments app. It uses Gemini models to analyze spending patterns, suggest savings actions, surface tailored offers, and explain financial concepts such as SIPs, credit scores, compounding interest, and tax regimes. Google says it can use transaction history and credit report data, accept text or voice queries, and respond in 10 Indian languages. 3

The availability statement is direct: the experience is launching in Google Pay in India, with the user opting in. Google also says the answers are for educational purposes and that financial data remains under the user's control. The announcement couples the AI launch with a new Google Pay Flex SBI Card, but that card is a separate credit product rather than a new model or API. 3

The product is a useful reminder that consumer AI expansion is moving through existing transactional surfaces. It also raises a higher bar for product testing: language coverage and convenience matter, but so do the provenance of personalized answers and the boundary between financial education and advice.

MCP changes the deployment assumptions

MCP 2026-07-28

The Model Context Protocol community officially released specification version 2026-07-28. The central change is the removal of protocol-level handshakes and session identifiers from the Streamable HTTP path. Each request carries its protocol version, client information, and capabilities, so a server can handle requests behind ordinary round-robin load balancing without sticky sessions or shared session storage. Developers that need application state must pass an explicit handle through tool arguments instead. 4

The release also moves method and tool names into HTTP headers for routing and metering, adds cache hints to list responses, and introduces Multi Round-Trip Requests for approval or input flows that previously depended on a held-open bidirectional stream. The revised authorization guidance hardens issuer validation and moves the ecosystem toward Client ID Metadata Documents. TypeScript, Python, Go, and C# Tier 1 SDKs were updated for the new version, with Rust support in beta. 4

The migration cost is real. Roots, Sampling, and Logging are deprecated, as is the legacy HTTP+SSE transport, with at least a twelve-month transition window. Existing implementations can continue to work, but new code should avoid building on the deprecated features. Teams with clients that depend on Mcp-Session-Id, logging/setLevel, or the old -32002 resource error code should audit those paths before advertising support for the new version. 4

Amazon Bedrock AgentCore Gateway

AWS published its implementation guide during the July 29 GMT+8 window and says AgentCore Gateway can support MCP 2026-07-28 alongside older versions. Adoption is an opt-in configuration change: administrators add the new version to supportedVersions, and clients select the version on each request. Older clients continue to use the earlier protocol version as long as the gateway still advertises it. 5

The practical warning is that UpdateGateway replaces the complete supported-version list rather than appending to it. An operator should read the current configuration first, send the full desired list, then test a request with the MCP-Protocol-Version: 2026-07-28 header. AWS says the gateway returns an HTTP 400 with code -32022 when a requested version is not advertised. 5

MCP's release is therefore more than a specification refresh. It changes where deployment complexity sits: less session management in the transport, more explicit state handles, version negotiation, and migration work in clients and gateways.

Security tools catch up with agent-generated software

CodeQL 2.26.1

GitHub released CodeQL 2.26.1 with expanded modeling for Go's log/slog, Apache POI sources and sinks in Java and Kotlin, Angular @HostListener message handlers in JavaScript and TypeScript, and fewer false positives in Rust's hard-coded cryptographic value query. It also updates several security queries, including path injection, SSRF, log injection, clear-text logging, and missing-origin checks. 6

The release is automatically deployed to users of GitHub code scanning on github.com. GitHub says the new functionality will enter a future GitHub Enterprise Server release, while older GHES installations can manually upgrade CodeQL where their deployment allows it. This is a maintenance release rather than a new AI model, but it matters to teams adding AI-generated code to repositories where false positives and missing framework models directly affect review load. 6

Immersive One Agentic Harness

Immersive Labs announced an Agentic Harness inside Immersive One to verify autonomous AI safety, control token spend, and measure secure and efficient AI operations. The company positions it for technical teams and leaders that need evidence about how autonomous systems behave, rather than another chat interface. The launch page currently directs prospective users to request a demo, so the status is announced and sales-led rather than self-serve. 7

The public announcement does not provide a benchmark, price, supported agent frameworks, or a detailed integration list. Those fields should be confirmed before a procurement decision. The product is still relevant to engineering leaders because it treats token spend and autonomous behavior as testable operational properties, not only model-quality metrics.

Launch-day tools for coding and site building

Prelint

Prelint launched on Product Hunt with a narrow claim: review every pull request against a team's architecture decision records, documentation, and past decisions, then catch product drift before merge. The product page says teams using several AI reviewers have seen about 40% of issues fixed before merge caught by Prelint, but that is a vendor claim without a public benchmark in the listing. 8

The listing shows free options and marks the product as launching today. The useful distinction from a conventional code reviewer is the reference set: Prelint is checking whether a change matches prior product decisions, not only whether the code passes tests or static analysis. Teams evaluating it should ask how it handles stale ADRs and conflicting documentation before trusting the percentage claim. 9

MemoryCustodian

MemoryCustodian launched as an open-source, local-first memory layer for coding agents. It stores decisions, constraints, rejected approaches, and project context as Markdown in the repository, then uses a manifest to load only the memory relevant to the current task. The Product Hunt listing names Codex, Claude Code, Gemini, and other coding agents as targets, and links to a GitHub repository rather than a hosted service. 10

That design makes the tradeoff explicit. The memory is reviewable, versioned, shareable, and deletable like code, but it also becomes part of the repository's governance and access model. The listing marks it free and launching today, so a developer can test the approach without adding a hosted memory dependency. 9

/mission for Claude Code

Spine launched /mission for Claude Code, a free plugin from its Medley project. Typing /mission turns an outcome into a live graph, coordinates Claude Code and Codex workers, lets the user review the result, and keeps the work going. The listing also says the plugin supports bring-your-own-key access through OpenRouter for models including Kimi and GLM. 11

The availability status is clear: the Product Hunt page says the plugin is launching today and free. The architecture is also easy to understand. It is a coordination layer around existing coding agents, not a new foundation model, so the main evaluation questions are task decomposition, worker isolation, review ergonomics, and the cost behavior of the BYOK route. 9

Framer AI Agents

Framer's July 29 Product Hunt launch puts agents on the design canvas and pairs them with branching and team publishing. Framer's product page describes the service as an AI website builder where users design with agents, refine on the canvas, and ship with a team. A launch-day comment from Framer says the new workflow can connect external AI tools such as Claude Code and Codex to Framer. 12 9

This is a product-surface change rather than a standalone model release. The test for teams is whether an agent can make useful changes without flattening the design system, and whether branching gives humans a practical rollback path when an AI edit goes in the wrong direction.

What to act on

Testing background agents: Gemini Spark is the item to watch, but its India rollout is staged and the post does not expose a public API or enterprise control plane. Start with low-risk read-and-draft workflows.
Maintaining MCP infrastructure: audit session assumptions, deprecated features, error-code handling, and version negotiation before enabling 2026-07-28 in production. Run old and new clients side by side where possible.
Reviewing AI-generated code: CodeQL 2.26.1 is already on GitHub code scanning. Prelint adds a product-decision layer, while Immersive One is positioned around autonomous-agent safety and spend. These address different failure modes.
Trying agent memory or orchestration: MemoryCustodian and /mission are both launch-day, low-friction experiments. The first stores context in the repo; the second coordinates multiple workers. Measure review time and repository noise before standardizing either.
Using AI inside customer workflows: Ask Google Pay is opt-in and India-only in this announcement. Its financial-data boundary and educational disclaimer should be part of any rollout assessment.

The July 29 window was defined by AI products becoming more persistent and more embedded in existing work surfaces. The engineering cost moves with that shift: background agents need permission boundaries, protocol upgrades need migration plans, and AI-generated software needs checks against both security rules and product intent.

Coverage is limited to public official announcements, release notes, and launch pages with a verifiable July 29 (GMT+8) publication or launch signal. Regional rollouts, unindexed changelog edits, and products without enough detail to establish capability and availability are not included.

AI Product Updates, July 29: Gemini Goes Proactive, MCP Goes Stateless, and Agent Tools Multiply