Top story

Google's Flash model now does last year's flagship work — and its agents go persistent

Source CNBC

Why it matters for entrepreneurs: At I/O 2026, Google made Gemini 3.5 Flash the default model behind the Gemini app and AI Mode in Search. Flash is the cheap, fast tier — $1.50 per million input tokens, $9 per million output — and it now beats last year's flagship Pro model on coding and agentic benchmarks. In plain terms: the budget tier crossed the line where it is good enough for real automation work, so the per-task cost of running an agent for a client just dropped. Be precise about "cheap," though. Gemini 3.5 Flash runs roughly 40% below Gemini 3.1 Pro but several times the price of the older Flash-Lite tiers. Cheaper than frontier, not cheapest on the menu.

Google also pushed its agents from "ask and answer" to "running in the background." Gemini Spark is a general-purpose agent that reasons across your connected apps; Daily Brief assembles a morning summary without being asked. Both are in beta, gated to AI Ultra subscribers and trusted testers first. The pattern is the real headline: the agent is becoming a standing feature of the app, not a tool you open. For a service business, the cost question shifts from "can we afford to run agents" to "which client workflows are now cheap enough to automate that were not three months ago." Re-run that math this quarter — the answer changed.

Quick hits

Anthropic lets agents reach inside your firewall — without opening it

Source InfoQ

At its Code with Claude event on May 19, Anthropic shipped two infrastructure features for Claude Managed Agents. Self-hosted sandboxes (public beta) run an agent's tool execution on infrastructure you control, or on providers like Cloudflare, Daytona, Modal, and Vercel. MCP tunnels (research preview) let an agent call internal databases, APIs, and ticketing systems through an outbound-only encrypted gateway — no inbound firewall holes. MCP, the Model Context Protocol, is the open standard that lets agents call outside tools. This is the unglamorous plumbing that gets an agent past a client's security review. If a client has refused to let an agent near internal data, this is the answer to point at.

Google is moving conversational ads inside AI Search answers

Source Google

Google is expanding ads inside its AI-generated Search results, with sponsored products appearing alongside the AI explanation and some ad units carrying their own built-in chatbot. For any service business that depends on being found, the storefront is moving: the AI answer is becoming the first screen a customer sees, and paid placement is moving into it. Check how your business shows up in an AI Search answer now, before the paid slots harden and organic mentions get crowded down the page.

Tool / launch watch

Two launches worth watching if you run an operations-heavy business or serve clients who do. Camunda opened a closed beta of ProcessOS on May 20 — an intelligence layer that maps an existing business process, then rebuilds it as a governed agentic workflow with human review built in. You describe the outcome in plain language; it generates the process, integrations, data mapping, and forms. It is aimed at teams running complex ERP and CRM stacks (Business Wire). Separately, Salesforce put Agentforce Coworker into beta — an agent that lives in the Salesforce search bar, pulls CRM context, and takes action, reachable from Slack, Teams, and ChatGPT (Salesforce Ben).

Neither is a finished product. But both confirm the same thing: if a client already runs Salesforce or a real process stack, the agent is arriving inside that tool whether or not they chose it. The deployable move this quarter is not buying a new agent platform — it is checking which agents the client's existing platforms are about to switch on, and getting their data and permissions ready before that happens.

Funding / M&A pulse

Angle for the blog

Headline: "Stop Shopping for an AI Agent. Your Client Already Bought Five." This week proved a shift a service firm can sell against. The agent layer is consolidating into incumbent software — Salesforce put one in the search bar, Camunda is rebuilding ERP processes as agentic workflows, Google made a capable agent the default in its app and its search box, and the budget-tier model that powers it all just got good enough for real work. The contrarian point for a 600-word post: most small businesses do not need to evaluate another agent startup. They are already paying for platforms that are switching agents on by default. The problem is that those agents are only as useful as the data and processes they can act on, and most clients' operations are not legible enough for an agent to touch safely — undocumented workflows, messy CRM records, permissions nobody has reviewed in two years. So the post reframes the SyncBroad AI offer from "we will build you an agent" to "we will make your operation ready for the agents you already own." Name the deliverable plainly: an agent-readiness audit. It maps which of the client's existing platforms ship native agents, scores their data hygiene and process documentation, scopes what an agent should and should not be allowed to do, and identifies the one or two workflows that genuinely cross platform boundaries — because that cross-platform plumbing (the MCP-tunnel layer Anthropic shipped this week) is the part no single vendor will do for them, and the part worth a custom build. The throughline ties straight to "Simplify It": simplifying a client's operations is no longer about adding an AI tool — it is about making the operation clean enough that the AI already in their stack can do the work without breaking something. Keep this a seed — leave the audit checklist and the pricing for the post itself.

The Tech Digest is compiled each morning by SyncBroad AI — a plain-English read on AI for service businesses. Browse the full archive, or book a 15-minute demo to see what's actually deployable for your operation.