Salesforce acquires Fin for $3.6 billion

Good morning 👋 Salesforce just paid $3.6 billion for an AI agent that closes 76% of support tickets on its own. That number is now the production benchmark.

In today's issue:

• 🔭 What the Fin acquisition says about the price of production AI agents
• 🧠 AgentCore Web Search goes GA
• 🔬 Bias propagates fast in multi-agent pipelines
• 🛠️ How to benchmark your agent before launch, plus five trending tools
• 🔗 Quick links

🔭 THE ONE THING

💰 Salesforce acquires Fin for $3.6 billion

Salesforce signed a definitive agreement on June 15 to buy Fin, the AI customer service agent that spun out of Intercom, for $3.6 billion. Fin resolves 76% of customer support volume end-to-end across chat, email, WhatsApp, SMS, phone, and Slack, without human escalation. That resolution rate is the whole argument for the price.

Agentforce was already at $1.2B ARR, up 205% year-over-year coming into this. Salesforce bought Fin anyway, for 30,000 customers and a packaged deployment model they chose to acquire rather than build. Deal closes Q4 FY2027.

For anyone building customer-facing agents: 76% autonomous resolution is the floor buyers are about to expect. Start measuring where you are.

🧠 MODELS & RELEASES

• 🌐 Amazon Bedrock AgentCore launched Web Search as a generally available feature, giving production agents live web access within the existing runtime. Already on AgentCore Harness? It plugs in without extra plumbing.

🔬 RESEARCH HIGHLIGHTS

• Contagion Networks: bias doesn't stay local in multi-agent systems. A new framework formalizes how evaluator biases spread when LLMs serve as judges inside multi-agent pipelines. A controlled three-agent experiment confirmed the propagation is fast and measurable. If you're using an LLM as a judge anywhere in your pipeline, audit it like the root of a tree, not a leaf.

• LedgerAgent: instead of restuffing task state into the prompt on every turn, this paper proposes typed ledger objects that carry facts, constraints, and domain conditions across a conversation. Cleaner than context stuffing, and auditable by design. Worth reading for anyone building multi-turn tool-calling flows with policy requirements.

🛠️ TOOLS & HOW-TO

Benchmark your agent before it touches customers (four steps)

The Fin acquisition sets a concrete bar. Here's a four-step eval to find out where you stand before launch:

1. Pull 200 real queries from your support logs. Weight toward high-frequency topics, not edge cases.

2. Run each through your agent. Log every escalation, every wrong answer, every hedge.

3. Score each as RESOLVED (correct and complete), PARTIAL (helped but needed follow-up), or ESCALATE (punted to a human).

4. Set your threshold before you look at results. If you pick it after, you'll rationalize whatever number comes back.

Prompt: Handle this support ticket: [TICKET]. Use any tools available. At the end, output one of: RESOLVED | PARTIAL | ESCALATE — then your customer-facing response, then one sentence explaining your classification.

5 trending right now

• 🎬 hyperframes: write HTML, render video, built for agents. HeyGen's open-source agent-native video pipeline. link
• 🧠 agentmemory: benchmarked persistent memory layer for AI coding agents. link
• 👁️ Crucix: personal intelligence agent that watches multiple data sources and pings you when something changes. link
• 📐 fireworks-tech-graph: production SVG/PNG technical diagrams from natural language, 7 styles including UML and agent workflow patterns. link
• 📚 ai-engineering-from-scratch: learn-build-ship curriculum for AI engineering, one of the fastest-growing repos in this space right now. link

🔗 QUICK LINKS

• NVIDIA at Cannes Lions AI advertising and marketing partnerships dominated the festival this week, with NVIDIA framing the shift from digital speed to AI autonomous operations for agencies. link
• Adobe + Amazon Quick Adobe Marketing Agent for Amazon Quick launched via MCP, pulling Adobe audience rankings and loyalty insights into Quick's interface without leaving the tool. link
• Anthropic Seoul Anthropic opened its third APAC office this week and signed an AI safety and cybersecurity MOU with Korea's Ministry of Science and ICT, the same week the US export ban cut Korean access to Fable 5 and Mythos 5. *(Primary Anthropic source unconfirmed; see footer.)*

See you tomorrow.

Pradeep Perugu