I/O 2026 ships Gemini 3.5 — and the rest of the agent stack moves in the same week.

This reframes the Teamwork Graph as a workforce knowledge graph rather than a work-only one — a substrate Rovo and Talent can both reason over. Competitors building AI on work data (Notion, Linear, monday.com) now need their own HR plumbing to keep up, and HR analytics vendors should expect Atlassian to encroach on their patch.

The agentic observability category just got its most coherent statement yet. Auto-investigation collapses the most expensive minute of incident response - context-gathering - and multiplayer Canvas makes the agent a teammate rather than a chatbot. Competitors selling 'AI assistants' bolted onto dashboards will be measured against an integrated investigation surface that connects alerts, telemetry, code, and chat. Expect Datadog, New Relic, and Grafana to react on roadmap if not on stage.

For developers this means more consistent results without learning model quirks; for the model vendors plugged into Copilot it means GitHub now controls demand allocation. A routing layer that ranks providers by health, cost, and task fit makes Copilot the load balancer of the IDE — and weakens any one model's ability to claim the developer relationship directly. Expect the picker UI to keep shrinking until manual selection becomes a power-user toggle.

This reframes the Informatica acquisition. By making it explicitly headless and cross-cloud — including MCP servers in AWS's Agent Registry — Salesforce is positioning Informatica to live inside competitor platforms, not just inside Data Cloud. The thesis is that enterprises will deploy agents across multiple clouds and need a consistent trusted-data layer beneath them; whoever owns that layer collects rent from every agentic workload regardless of where the agent runs. This puts direct pressure on Databricks, Snowflake, and SAP's data fabric narrative, and signals that Salesforce sees the agentic-AI infrastructure layer as a bigger prize than the CRM-adjacent agent itself.

The interesting move is monetization, not technology — Zoho is converting analytics demand inside its CRM base into Zoho Analytics revenue, instead of letting it leak to Tableau, Power BI, or vertical BI tools. Expect the same playbook against Books, Desk, and Inventory next. For embedded analytics vendors selling into Zoho CRM customers, the addressable market just got smaller.

Form generation is the natural beachhead for LLM-in-ESP because forms are bounded, visual, and easy to evaluate — if AWeber can deliver here, the next obvious step is the email editor itself. It also raises the bar for Mailchimp, ConvertKit, and Beehiiv to ship comparable generation tools or cede the small-business segment's perception of being 'AI-native.' Watch which competitor moves first.

Nested virtualization is the long-standing missing piece that keeps mobile teams and infra-heavy test pipelines tied to self-hosted runners. Shipping it as default-on, no-extra-cost narrows Depot's gap with self-hosting and opens two markets that have been awkward on managed CI for years: Android development and Kubernetes/infra testing. It's the first release in this batch that competitors can't match with a UI update.

Text-to-SQL has been generally accepted as fragile precisely because users phrase ambiguous questions and BI tools answer them confidently anyway. By forcing a back-and-forth, Holistics is conceding that the right answer often requires negotiation — a stance that, if it works, raises the bar for competitors still shipping single-turn Ask-anything boxes. Expect more BI vendors to follow with their own clarification loops within a quarter.

For enterprise buyers locked into specific provider relationships, a new Cohere flagship resets the 'is Cohere still competitive on Tier-1 reasoning?' question that gets asked at every renewal. It also tightens the case for sticking with the Command line through deprecation of the old r/r-plus generation. Watch for AWS Bedrock and Oracle to update their Cohere-on-cloud SKUs within a release cycle, and for fine-tuning customers to push for A+ tuning options that aren't yet visible.

If Gemini 3.5's action layer actually works in third-party hands, the competitive frame shifts from raw model benchmarks to agent reliability and tool integration — a surface where Google has Workspace and Search distribution advantages OpenAI and Anthropic don't. It also justifies the $100 AI Ultra tier and gives Workspace customers a reason not to wait out enterprise agent rollouts elsewhere. Watch for Microsoft to respond on the Copilot agent surface within the quarter.

Always-on, action-taking consumer AI is the part of the agentic story most exposed to trust failures — an action that misfires in a chat is forgettable; one that misfires in a 24/7 background agent is reputational. By committing this surface to Gemini 3.5's action capabilities, Google is making a public bet that the model is good enough to operate this way at scale. Competitors who currently ship narrower assistants will need to decide whether to match the posture or differentiate on bounded reliability.

Changelog tools have spent two years bolting AI suggestions onto existing editors. Smart Draft inverts that — the editor is now the polishing surface, with AI as the authoring surface. For product marketing teams whose changelogs need consistent voice across many contributors, that's the actual unsolved problem. Competitors (Beamer, Headway, Releasenotes.io) without a brand-voice-aware multi-input drafting flow now look one product cycle behind.

Agency operators spend most of their time wiring If/Else logic in workflow builders — it's the labor that justifies HighLevel's per-agency pricing. Letting the agent do branching natively undercuts that labor and pushes complexity into a single configurable surface. If the pattern holds, expect competitors (GoHighLevel-style agency CRMs and adjacent marketing-automation tools) to face the same pressure: their workflow builders look manual next to an agent that just decides.

Cash application is the textbook back-office task to hand to an agent: high volume, repetitive, mostly unambiguous, expensive in headcount. By scoping the agent to clear matches only and exposing per-transaction undo, Upflow gets to charge for autonomy without taking on the audit risk that has slowed other AR vendors. Expect this pattern — admin toggle, conservative defaults, visible auto-apply status, easy undo — to become Upflow's house style for shipping further agents, and pressure on the BlackLine / HighRadius end of the market where cash application has historically been the headline use case.

CRM AI is high-stakes — a hallucinated quote inserted into a deal record can torpedo a relationship. Citations turn AI from a black box into a tool reps can defend in front of a manager or a deal team, which is the only way AI features reach durable adoption in revenue org. Expect other CRMs (Salesforce Einstein, HubSpot Breeze, Attio AI) to follow with similar source-traceable output; the bar for shipping AI in revenue tooling without provenance is rising fast.

Marketing automation has been a soft target for AI assistants because the workflows are routine and the data shapes are well-defined. Mailchimp, Klaviyo, and ActiveCampaign do not have shipped MCP servers yet — by being first in the creator-economy slice, Kit makes it easier for AI tools to recommend it as the default integration. If MCP becomes the way AI assistants talk to marketing tools, being the well-known reference implementation in a category compounds quickly.

Pricing has been Vercel's single biggest narrative risk; the typical anti-Vercel post hinges on a $40k bandwidth invoice rather than the product itself. If Flat Rate works, the conversion story to Pro tightens and Cloudflare Pages, Netlify, and Render lose their main rhetorical opening. The format Vercel picks here — what the cap covers, what triggers overage — will set the bar competitors have to match or beat.

Public price/perf comparisons against named frontier APIs are the playbook for neoclouds pursuing high-volume agent workloads, and coding agents are the highest-token-burn category right now. If the numbers hold up in practice, this puts real cost pressure on closed-lab agents at the same time Anthropic and OpenAI are trying to raise prices on long-running tasks. Expect competing inference vendors to publish their own coding-agent benchmarks within weeks.

Email-marketing tools are notoriously menu-heavy; creators rarely use more than 20% of the surface. Wiring Kit into an AI assistant collapses that complexity into 'just describe what you want.' Mailchimp, Beehiiv, and ConvertKit competitors don't have equivalent MCP coverage at this depth. If the beta holds, Kit's pitch shifts from 'best automation builder' to 'the email backend your AI uses,' which is a more defensible position as agent usage rises.

This is a clear scope expansion from AP automation into procure-to-pay. For DACH SMB and mid-market finance teams, the alternative has been either Excel-based requisitions or stitching a separate procurement tool to an AP automation product. By owning both sides of the match, Candis can charge for the wider workflow and lock in customers who would otherwise consider standalone procurement vendors like Pleo's Procurement, Penny, or Mercanis. Expect competitors in the German AP space to react — DATEV's own roadmap especially, since this leans on DATEV-native primitives.