Signal Brief #6: Amplitude’s Code Mode + data-quality agent + 4 more

SIGNAL BRIEF

Issue #6 · Monday, August 17, 2026 · 5 AI tools for B2B SaaS product managers

Amplitude introduced two AI capabilities on Wednesday aimed at the question its launch post asks directly: AI can build — but can it know what worked? Code Mode brings code-level analysis inside Amplitude itself: teams can run SQL and Python against the behavioral data already in the platform to investigate questions that go beyond a standard chart — what changed, why it matters, who is affected, and what to do next — without exporting data, filing a ticket to a data scientist, or waiting on a separate analysis cycle. It is available to all Free, Plus, and Growth customers. Data Assistant Agent attacks the other half of the problem: the trustworthiness of the data the reasoning runs on. It reviews taxonomy, metadata, usage patterns, and project context to find unclear, duplicated, or undefined events and properties, explains why each issue matters, and recommends prioritized fixes — and it can run an AI-specific audit, investigate taxonomy issues, explain its own prioritization, and draft metadata improvements, all in chat. It is available now to Plus, Growth, and Enterprise customers with AI features enabled.

Why it matters for PMs: Every AI feature you ship is about to be judged on two things — how deeply it can reason, and whether the data underneath is trustworthy enough to reason from — and Amplitude is now selling both at once. Code Mode collapses the distance between “the chart looks wrong” and a code-level answer, while Data Assistant Agent treats data quality as an agent problem, not just a human chore — because agents will read your taxonomy whether you clean it up or not. If you are instrumenting your product for AI features, the audit is worth running before the agents arrive.

Google released Gemini 3.7 Flash on Wednesday, calling it its most intelligent workhorse model yet for coding and agents — and it arrives just three weeks after 3.6 Flash, a cadence Google attributes to developer feedback and algorithmic innovations. The gains are aimed squarely at production work: substantial improvements across software engineering, knowledge work, and web development workflows, with better debugging, more deployable production-ready code on the first try, and completed apps in fewer prompts. The headline number is price: an introductory rate of half the original 3.6 Flash cost per million tokens. Gemini Spark — the 24/7 personal agent for Google AI Pro and Ultra subscribers in more than 160 countries — moves to 3.7 Flash immediately, with improved tool use across Google Workspace apps for complex, multi-skill workflows. Developers can access the model in Google Antigravity, the Gemini API, Google AI Studio, and Android Studio; enterprises get it in the Gemini Enterprise Agent Platform and the Gemini Enterprise app. Bloomberg notes the release lands while Gemini 3.5 Pro, the top-tier model, remains delayed.

Why it matters for PMs: The model layer under your AI features just got cheaper and more capable at the same moment — half-price introductory pricing on a workhorse-class model changes the unit economics of every per-seat AI feature on your roadmap. The cadence is the other story: three weeks between Flash generations means model selection is a moving target, and “which model, at what effort, at what price” is now a product decision with roadmap consequences, not an infrastructure afterthought.

Figma’s release notes on Wednesday added a discovery, creation, and sharing layer for the Figma agent: an AI skills library in the Figma Community with more than 50 skills for design workflows — research, design systems, and handoff among them. You can add a skill straight to your agent, or copy its markdown file into the agent of your choice. Creation is built in too: ask the Figma agent to make a skill from the context in a file and it packages a repeatable workflow into a slash command you can invoke in the agent; publish it to the Community and anyone can try it or remix it for their own workflows. The release extends the skills model Figma introduced at Config — skills as portable, shareable units of how a team works — and lands alongside last week’s update letting agents run Figma Weave tools from ChatGPT, Claude, or Cursor through the Figma MCP server.

Why it matters for PMs: Design process is becoming a distributable artifact. If your team works in Figma, the community library is a shortcut to design-system audits, research synthesis, and handoff checks — and the create-with-the-agent flow means designers can encode their own conventions without writing code. The pattern is worth stealing too: a marketplace where power users package their workflows as skills is a retention and distribution engine, whatever your product category.

Anthropic announced Thursday that Claude Tag — the feature that adds Claude to a Slack channel to work alongside your team — now uses context from across the channel, plus its memory and the standing instructions you have given it, to decide when to contribute. With the old classifier removed, Claude picks one of four moves for any message: reply inline when the answer is short, verifiable, and something the channel does not already know; start deeper work in a thread when a message deserves real time; route the message into a workstream it already has open; or say nothing when nothing is called for. Anthropic says the update makes Claude roughly 30 percent better at judging when — and when not — to respond proactively, and faster to respond as well; the additional context does not count toward usage or spend limits on any plan. The update is live today for Claude Teams and Enterprise customers, at no additional cost.

Why it matters for PMs: “An annoying agent is worse than an unhelpful one” — Anthropic’s framing names the design bar for every agent that lands in a shared surface, and the four-move taxonomy (reply, thread, route, stay silent) is a spec worth stealing for anything collaborative you ship. The deeper lesson: silence is a designed outcome. If your roadmap includes agents that act in team spaces, judge them on when they don’t speak as much as when they do.

Fin — the customer-agent company formerly known as Intercom — shipped memory on Wednesday: the Fin AI agent now remembers past conversations, so returning customers pick up at the next step instead of starting over. The memory is not tied to a channel — the same context travels with the customer from chat to email, keeping the experience joined up however they get in touch — and Fin draws on what a customer has shared before so answers feel like they come from a company that actually knows them. It is the latest step in a fast run for Fin: Operator, its AI agent for support admins that reads conversations, flags why CSAT moved, and drafts fixes as before-and-after diffs for human sign-off, went generally available earlier this month, and Fin reports customers have used it to make more than 20,000 support improvements in three months.

Why it matters for PMs: Cross-channel memory is becoming table stakes for customer-facing AI — and in B2B SaaS, support is part of the product. If the agent on your site asks a customer to repeat themselves, that is now a product defect, not a vendor limitation. Memory also raises the bar for what “knowing the customer” means across the stack: the same expectation is heading for onboarding, success, and sales surfaces. And note the rebrand — Intercom renaming itself around its agent says where the company believes the value has moved.