DoorDash's agent lifted checkout 24%. Here's how.

In this edition:

• This week: DoorDash published 24% conversion lifts from its deployed agent, Mira Murati's Thinking Machines released its first open-weights model, and Prefect acquired Dagster to build agentic orchestration

• Under the radar: DeepSeek is closing in on $500M in annualized revenue and weighing an IPO

• What's on the calendar: Anthropic's credit line discussions, Prefect/Dagster unified product timeline, and AI4 2026 in Las Vegas August 11-13

THE WEEK IN AI

THE WEEK IN ONE SENTENCE

The operators who published proof this week (actual conversion lifts, open weights, governance scaffolding for autonomous agents) are setting the standard everyone else is going to be measured against.

THREE SIGNALS

01 • Agents

DoorDash published the numbers on its AI agent, and they're specific enough to matter.

DoorDash spent the better part of a year building Ask DoorDash, an AI shopping assistant that combines large language models with specialized agents, MCP tooling, and a memory layer that tracks what a customer has ordered, saved, and browsed. Last week, they published the architecture and the results: up to 24% higher checkout conversion and 17% larger basket sizes.

Those numbers are not remarkable by themselves. What's notable is that DoorDash chose to publish them. The company built a measurable outcome into the design from the start, ran the agent against that outcome, and then shared what happened. That's a different posture from most companies, which deploy agents, declare them useful, and move on.

The architecture is also worth examining. Most enterprise AI right now runs a single large model and calls it an agent. Ask DoorDash uses a tiered structure: a generalist model handles language, specialized agents handle specific tasks, MCP tools handle system access, and a memory layer handles context across sessions. It's closer to how a well-designed software system works than how most "AI assistant" products are built. The operator question this raises is not whether to build something like this. It's whether you have a clear measure today for what you'd need to see before you'd scale it.

TEAM READINESS AUDIT

Get your team’s AI readiness score. Leave with an AI readiness brief.

A 15-minute assessment that turns your answers into a real decision artifact: your readiness level, your six-axis shape, where you're strongest, where you're constrained, and what not to build yet.

02 • Models

Thinking Machines Lab shipped its first model, and Mira Murati made a deliberate choice about trust.

Mira Murati left OpenAI in September 2024 as its CTO. On July 15, her new company released Inkling: a 975 billion parameter Mixture-of-Experts model, 41 billion active, with full open weights. It handles text, images, and audio, supports a 1 million token context window, and is available today for fine-tuning through their Tinker platform.

The model is not the strongest available. Murati said that explicitly: "Inkling is not the strongest overall model available today, open or closed." What it is, she argued, is a good base to customize, broad enough to adapt, transparent enough to inspect.

The decision to release full weights is not a technical one. It's a positioning choice. At the moment when OpenAI and Anthropic are building toward closed deployment deals with governments and enterprises, Thinking Machines is betting that some significant portion of the market will want to inspect what they're running before they trust it. Whether that bet lands depends on what Inkling actually does inside the companies that fine-tune it. What's clearer is that the market is splitting between operators who will accept frontier capability on faith and operators who need to open the hood before they commit, and Thinking Machines is explicitly going after the second group.

03 • Infrastructure

Prefect acquired Dagster, and the reason they gave for doing it is the more interesting part.

Prefect and Dagster spent years as direct competitors in the workflow orchestration space. Both were built for teams managing complex data pipelines: scheduling, retries, observability, and lineage. Prefect CEO Jeff Lawson announced the acquisition this week, saying the combined company would keep both products running independently and turn toward "agentic orchestration."

The framing is specific. Lawson described the problem as: how do you automate software that is itself autonomous? The answer he's building toward combines Dagster's declarative approach (what should be achieved) with Prefect's durable execution model (following paths that can't be known in advance), plus FastMCP for governed tool access.

For operators, the consolidation is a signal worth watching. The companies quietly winning the infrastructure layer right now are the ones building the scaffolding that goes around the AI, not just the AI itself. Scheduling, governance, observability, retry logic are not interesting to talk about, and they're what makes agents reliable enough to run in production. Prefect's bet is that agentic orchestration will be a larger market than data orchestration, and they are positioning for it before most of their future customers know they'll need it.

UNDER THE RADAR

DeepSeek, the Chinese AI lab that rattled markets in January with its low-cost reasoning model, is closing in on $500 million in annualized revenue and is weighing an IPO, according to reporting this week. The figure matters for context more than the headline.

Most Western coverage of DeepSeek focuses on the model releases. The revenue number is the more durable fact. A Chinese AI lab that demonstrated frontier reasoning capability at a fraction of the cost, and then turned that capability into half a billion dollars in annual revenue, is not an academic curiosity. It's a competitor. DeepSeek doesn't sell to most of the operators reading this. But it is setting the floor on what inference should cost, and every vendor pricing their API above that floor is going to have a harder time defending the gap.

What I keep hearing in conversations with operators is the assumption that DeepSeek is "too risky" for enterprise use and therefore not relevant to their buying decisions. The risk question is real. The relevance question is separate. DeepSeek's economics are reshaping what's negotiable, even for companies that will never deploy a Chinese model.

QUOTE OF THE WEEK

"Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning."

This may be the first major model announcement in two years to lead with what the model is not.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

WHAT’S ON THE CALENDAR

• Anthropic is in discussions to add several billion to its bank credit line, ahead of what's expected to be a major capital raise. No close date reported.

• Prefect and Dagster's combined orchestration stack has no shipped timeline yet. Both open-source communities are watching for when the unified product arrives.

• AI4 2026 runs August 11-13 in Las Vegas. The largest US conference for the enterprise AI operator audience this summer, with a dedicated deployment track.

P.S. If you're heading into Monday without a clear way to measure whether your AI deployment is actually working, that's the question DoorDash answered this week. Hit reply and tell me: what's the one outcome you'd need to see before you'd scale what you're running? I read every reply.

Have a good weekend,
Haroon