Research / Weekly Deep Dive
How agents actually ship
One researched edition a week. Real deployments with measured outcomes, the frameworks and harnesses behind them, and the honest counter-facts, sourced from primary research, not hype blogs.
Edition · August 3, 2026
Advantage Solutions redirects 70,000 hours a year by taking agents past the pilot
A sales-and-marketing services firm put Claude Enterprise into three live workflows and redirected 70,000+ labor hours a year, with governance, not tooling, as the deciding factor.
Caylent builds its own agent orchestrator and cuts a 2,600-hour migration to 220
Instead of buying a migration tool, Caylent built a multi-agent orchestrator on the Claude Agent SDK that runs inside the customer's own cloud and cut one engagement from 2,600 engineering hours to about 220.
Eve makes the buy-a-vertical-vendor case for small plaintiff law firms
Plaintiff firms with no in-house ML team buy their agents off the shelf, and Eve customers report large jumps in lead conversion, case value, and drafting speed.
Suzano puts a natural-language-to-SQL agent in front of 50,000 staff
A Brazilian pulp-and-paper maker built a Gemini agent that turns plain-language questions into SQL, which Google Cloud says cut query time by about 95 percent for 50,000 employees.
Edition · July 27, 2026
The Clinician cuts clinical project setup from 42 days to about 3 hours
A health-tech platform built a no-code authoring wizard on Gemini and took new-project setup from a median of 42 days to 3.1 hours.
Lyft builds a self-serve platform so non-engineers ship the agents
Lyft built an internal platform on LangGraph and LangSmith that lets non-technical support experts create and tune agents, cutting agent build time from about six months to a few weeks.
OpenAI dogfoods its own managed agent product on live phone support
Alongside launching Presence, its enterprise agent product, OpenAI ran it on its own phone support line and reports resolving 75 percent of inbound issues without a human.
Mars puts a single governed model layer in front of 62,000 staff
Mars rolled out Gemini Enterprise as a shared, governed AI layer across its workforce, reporting campaign-creation timelines cut from months to days.
Edition · July 20, 2026
Fountain turns high-volume frontline hiring into an agent workflow
Fountain built an AI Recruiter that auto-screens candidates and runs voice interviews on Claude, cutting manual screening roughly in half and doubling conversion for employers like rideshare firm Alto.
Doctolib ships features about 40% faster after rolling out Claude Code
Europe healthcare booking platform Doctolib piloted Claude Code with 30 engineers, then went team-wide, rebuilding a legacy visual-regression test suite in hours and citing roughly 40% faster feature shipping.
eSentire compresses expert threat analysis from five hours to seven minutes
Managed security firm eSentire built an investigation agent on the Claude API that takes threat analysis from about five hours to seven minutes and agrees with senior human analysts around 95% of the time.
PwC builds on the open CrewAI framework instead of buying off the shelf
PwC used the open CrewAI framework to orchestrate agent crews for internal GenAI work, reporting code-generation accuracy rising from around 10% to over 70% and faster document turnaround.
Edition · July 13, 2026
Box goes AI native by building its agent on an open framework
Box rebuilt its enterprise content agent on LangGraph Deep Agents, trading hardcoded specialist sub agents for a parent agent that spawns children on demand, and says it iterates roughly 3x faster as a result.
TELUS puts a governed model layer in front of 57,000 staff
TELUS runs Fuel iX, an internal platform on Google Cloud that gives its workforce access to Claude and other models, and reports 500,000+ hours saved and more than $90M in business benefits.
Zapier runs 800+ narrow single purpose agents instead of one big one
Zapier has deployed more than 800 internal AI agents, most of them narrow single task bots, on top of 90%+ internal AI adoption, a shape that is unusually copyable for a small business.
Edition · July 6, 2026
AES cuts a 14-day safety audit to one hour with a three-layer agent system
The power company put roughly 1,500 annual safety audits onto Claude agents via Google Cloud, keeping a human in the loop, and reports audit time falling from 14 days to one hour.
Uber saves about 21,000 developer hours generating tests on LangGraph
Uber built AutoCover, a multi-agent test generator wrapping LangGraph in an internal framework, and reports roughly 21,000 developer hours saved across about 5,000 engineers.
Danfoss automates 80% of B2B order decisions from plain email
Rather than build, Danfoss bought a specialist vendor whose agents read unstructured order emails and trigger ERP actions, reporting up to 80% of transactional decisions handled automatically.
Mid-market consumer brands buy their way to roughly 70% support resolution
Brands including OluKai and Madison Reed deployed managed customer-experience agents from Sierra rather than building, with vendor-cited resolution rates near 70% and a 50% drop in cancellations.
Edition · June 29, 2026
Gelato turns carrier onboarding from 5 days into 10 minutes with a CrewAI agent crew
The on-demand manufacturing platform built CrewAI agents that write, test and deploy carrier-integration code, cutting onboarding from 5 days to about 10 minutes. SKU-mapping timelines dropped more than 90%.
A small design shop scaled from 3 to 14 people on a CrewAI lead-generation crew
Brickell Digital replaced referral-only growth with CrewAI agents that research prospects, score them and prep sales calls. The firm reports qualified leads up more than 80%.
Tines collapses a 120-step security workflow into a single agent step on Claude via Amazon Bedrock
The automation platform rebuilt a 120-step playbook as one dynamic agent action, with the agent resolving branching logic at runtime. Tines describes roughly a 100x time-to-value improvement.
The LangGraph supervisor pattern keeps recurring for multilingual support
A coordinator agent that routes each conversation to narrow specialist sub-agents by language and product is the recurring LangGraph shape for support. Rakuten is reported to have cut average handling time about 40%.
Edition · June 15, 2026
AppFolio gives property managers back 10+ hours a week with a LangGraph copilot
AppFolio built Realm-X, a conversational agent for property management, on LangGraph and LangSmith, and reports users save 10 to 12 hours per week.
Postman runs Agent Mode on Claude via Amazon Bedrock
Postman built Agent Mode on Claude Opus 4.6 served through Amazon Bedrock, reporting up to 1,150 hours saved per developer per year.
Supermetrics automates the weekly marketing report on Google ADK
Supermetrics built a Marketing Intelligence Agent on Google ADK and Vertex AI Agent Builder, reporting 15+ hours freed per marketer each month.
Edition · June 8, 2026
KPMG rolls agents into real audit and supply-chain work across 276,000 staff
Announced June 9, KPMG is deploying AI agents firm-wide on Microsoft Agent 365 (lifecycle + governance) with Microsoft 365 Copilot, and the named use cases are concrete: agents embedded in Integra LifeSciences’ supply chain, regulatory and medical affairs, and in audit work for earlier risk identification and real-time analysis. A genuine deployment, not a pilot — built on KPMG Workbench atop Azure AI Foundry.
Podium cuts engineering escalations 90% with an agent on the support line
Podium built an agentic support system that, per LangChain’s customer write-up, reduced engineering intervention by 90% — routine support issues that used to pull engineers off product work are now resolved by the agent. Built on LangSmith for tracing and evaluation, it’s a clean example of the build path paying back in reclaimed engineering time.
Monday.com gets 9x faster feedback loops by instrumenting its agents
Monday.com reports (via LangChain) 9x faster feedback loops after building evaluation and feedback automation on LangSmith — the unglamorous but decisive work of measuring and iterating on agent quality, sped up nearly an order of magnitude. The takeaway: the teams that ship reliable agents are the ones that made iteration cheap.
Trellix takes security log parsing from days to minutes with a LangGraph agent
Security vendor Trellix built an agent that, per LangChain, cuts log parsing time from days to minutes — a concrete win in a domain where analyst hours are expensive and slow triage has real cost. Built on LangGraph for the orchestration and LangSmith for observability.
A startup compresses its internal tooling by building on agents instead of buying
In a June 5 Anthropic customer story, startup delight.ai describes building internal tools for its AI/ML teams with Claude Code — the small-team pattern of using a coding agent to absorb in-house tooling work rather than standing up or buying a separate platform. No published metrics yet, but the clearest SMB-shaped use case of the week.