Field Notes
Notes from the teambuilding production agentic AI.
September 30, 2026
What agent inference actually costs in production
The pilot cost forty dollars a month. The rollout cost eleven thousand. Here is where the money goes and which controls actually move the number.
September 30, 2026
Design the handoff before you design the agent
The escalation path is the product. Teams that build it last end up with an agent customers learn to fight their way past.
September 30, 2026
Test your agent before it talks to a customer
Most teams ship agents with less testing than they would give a checkout form. Here is the eval harness that catches the failures worth catching.
September 30, 2026
Deploying Claude on Bedrock with Data Cloud row-level security
FinTech architects need tenant isolation enforced at the data layer, not just IAM policies. Here's the production pattern for Claude on Bedrock with Data Cloud row-level access policies.
June 8, 2026
Anthropic's Sales Team Ships Cowork on Their Own Stack
Anthropic's Mid-Market GTM lead runs 4,000 accounts on Salesforce + BigQuery + Cowork. The pattern: AI augments ops rhythms, humans still own every customer send.
June 8, 2026
Salesforce shipped 151% more code with AI. Here's the architecture.
Removing token limits and standardizing on Claude Code took Salesforce from 90% adoption to 18x sprint compression. The workflow pattern they built is now the blueprint for mid-market multi-agent stacks.
June 8, 2026
Agentforce $1B ARR: What 28 Trillion Tokens Tell You About Production Deployment
Salesforce shipped 28.6 trillion tokens in Q1. That's not a demo metric — it's the architecture signal mid-market teams have been waiting for.
June 8, 2026
Opus 4.8 Delivers the Agent Reliability Lift Mid-Market Needed
Anthropic just shipped the first model to complete every case on a real Super-Agent benchmark. If you're running Agentforce or AgentCore workloads in production, this changes your error budget math this quarter.
June 8, 2026
TD Bank Just Shipped the Mortgage Agent Every Board Will Ask About
A Big Six bank put an agentic AI into live mortgage pre-adjudication under named governance, ten weeks before EU AI Act enforcement. If you're in regulated finance and haven't walked your board through this case, you will.
May 15, 2026
Mid-Market Teams Should Skip the LLM Gateway for Claude Code
If you're under 500 devs, adding a gateway between your engineers and Bedrock Claude Code is architectural theater that burns weeks you don't have.
May 15, 2026
You Can't Stop Developers Using Claude, So Route It Through Bedrock
Your devs are already pasting schemas into claude.ai. The answer isn't a Slack policy — it's Bedrock with PrivateLink so Claude never leaves your VPC.
May 15, 2026
AWS AgentCore Payments Ships Payment Rails Nobody Asked For
AWS just launched managed payment infrastructure for autonomous agents. Mid-market shops running Agentforce + Claude should ignore it for 18 months.
May 15, 2026
Why Your Cortex Code Budget Exploded This Month
Snowflake just published the governance script mid-market teams should have deployed six months ago. If you're running Cortex Code without RBAC + daily credit caps, your CFO already knows.
May 15, 2026
Databricks MCP Marketplace Fixes the Stale-Data Problem Agentforce Can't
Agentforce agents hallucinate less when they can check live credit scores. Databricks just shipped the plumbing to make that happen.
May 15, 2026
Stop Choosing Between Anthropic Direct and Bedrock Claude
The right answer is both. Route primary to Anthropic, failover to Bedrock when you hit 529s. That's the pattern shipping production Agentforce stacks this quarter.
May 15, 2026
Cortex Agents Will Leak Your Tenants' Data Without Row-Level Policy
Prompt injection is not the main risk. The main risk is that Snowflake Cortex Agents has no native tenant boundary, and your LLM will happily write SELECT * FROM customers WHERE 1=1.
May 13, 2026
Anthropic's SMB Play Exposes the Mid-Market Integration Tax
Anthropic's QuickBooks toggle is easy; wiring Claude API + Data Cloud + Unity Catalog for a 400-seat FinTech is not. The gap is where mid-market revenue dies.
May 13, 2026
LLM Fallback Is Infrastructure Now, Not a Nice-to-Have
If your production AI stack goes down when one LLM provider hiccups, you've built a house on sand. Fallback routing belongs in your orchestration layer, not your error handler.
May 13, 2026
The Salesforce-Anthropic integration: what it actually changes for mid-market architecture
The trust-boundary unlock matters more than the model upgrade. Here is the shift, and the new pattern most teams have not built yet.