How Apollo Uses Deep Agents and LangSmith for GTM AI
Apollo rebuilt its AI assistant on Deep Agents and LangSmith to run the full go-to-market loop — prospecting, enrichment, outreach, and analytics — through one agent with MCP integrations.
AI Daily Recap
18 articles · 5 categories
The finishable daily brief
What happened in AI — Jul 21, 2026
Tuesday, Jul 21, 2026
18 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Anthropic detailed how it secures an SDLC where AI now authors 80% of merged code; Datadog described a parallel pattern of having an agent write specs for a deterministic code-generation kernel.
- OpenAI and Hugging Face jointly disclosed a security incident that happened during AI model evaluation.
- Google Cloud moved CodeMender, its automated vulnerability-remediation agent, into preview.
- Moonshot AI's open-weight Kimi K3 is undercutting closed-source pricing, fueling $50B pre-IPO valuation talk and a public IP-theft accusation from the US Treasury Secretary.
- Google DeepMind shipped three new Gemini 3.x Flash model variants, and Cognition's Devin now runs its coding-agent work inside Modal sandboxes.
Today's engineering story is agents becoming production infrastructure, not demos: Anthropic and Datadog both published how they run AI-authored code at scale, while Cognition's Devin and Apollo's support stack show agent orchestration hardening around real workloads instead of prototypes.
A second thread is the Chinese open-weight price war — Moonshot's 2.8T-parameter Kimi K3 is undercutting closed-source pricing enough to fuel $50B pre-IPO valuation talk, and it drew a direct IP-theft jab from the US Treasury Secretary.
Agents in Production 3 items
Three separate teams described moving agent orchestration from prototype to production workload this week, each solving the same problem: keeping an autonomous agent's output reliable at scale.
How Datadog built a "universal machine tool" for Claude Code
Datadog has an agent write specifications that feed a deterministic kernel, which then generates the application code — trading direct LLM code generation for a spec-first, reproducible pipeline.
Devin Outposts on Modal
Cognition's Devin can now execute its coding-agent work inside Modal sandboxes via a new "Outposts" integration, moving agent execution off Devin's own infrastructure.
Observability & Feedback for Coding Agents 4 items
New tooling this week targets a specific gap: giving builders visibility and a feedback channel into what a coding or voice agent actually did, not just its final output.
Trace voice agents in LangSmith
LangSmith now traces voice agents built on Pipecat, LiveKit, OpenAI Realtime, and Gemini Live, capturing audio, STT/TTS latency, interruptions, and tool calls in one trace.
Show HN: Pinpoint – Visual feedback for AI coding agents
Pinpoint lets a developer drop contextual visual feedback directly onto an AI coding agent's UI output instead of typing it out, aimed at keeping the agent autonomous while still steerable.
Show HN: Observability for Coding Agents and LLM Applications
A new open tool adds tracing and observability specifically scoped to coding agents and LLM applications, an area general-purpose APM tools don't cover well.
How to build interactive experiences with canvases
GitHub Copilot's canvases turn agent output into an interactive workspace for visualizing information and taking action, rather than a static chat transcript.
Models & Infra 5 items
Model and infrastructure news split between new frontier releases and the machinery to run models at either end of the scale spectrum — gigascale training clusters and a local Mac laptop.
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
Google DeepMind released three new Gemini Flash-tier models in one announcement, expanding the fast/cheap end of its model lineup.
Kimi K3 rocks the AI industry as Moonshot AI undercuts closed-source American competitors on price
Moonshot's Kimi K3 is a 2.8T-parameter open-weight model priced well below closed-source US competitors, though it still needs serious hardware to deploy at scale.
Built for Vera Rubin, NVIDIA Spectrum-6 Arrives in Gigascale AI Factories
NVIDIA's Spectrum-6 networking is now shipping into AI data centers built around hundreds of thousands of GPUs and CPUs, the interconnect layer behind gigascale frontier training runs.
Nativ: Run AI models locally on your Mac
Nativ, from the developer behind MLX-VLM, wraps MLX to run local vision-LLMs on a Mac — the opposite end of the deployment spectrum from gigascale training clusters.
Yelp Unifies ML Model Training with Training Orchestrator
Yelp replaced its per-team Spark training scripts with a single configuration-driven, DAG-based Training Orchestrator, a common MLOps consolidation pattern as model counts grow.
Securing the AI-Native SDLC 3 items
As AI writes more of the code that ships, three items today addressed the same question from different angles: how do you secure a development process where the author is no longer only human.
How Anthropic secures its AI-native software development lifecycle
Anthropic's Deputy CISO, Jason Clinton, details how the Security Engineering team secures an SDLC where AI now authors 80% of merged code.
OpenAI and Hugging Face partner to address security incident during model evaluation
OpenAI and Hugging Face shared early findings on a security incident that occurred during AI model evaluation, framed as advanced cyber capability rather than routine misconfiguration.
Now in preview: Find and fix software vulnerabilities with CodeMender
Google Cloud's CodeMender, an agent that automates code vulnerability remediation, moved into preview as adversarial AI-driven attacks on code accelerate.
Chinese Open-Weight Labs: Money and Politics 3 items
Moonshot AI's pricing and valuation news collided with a Washington-level dispute over whether Chinese open-weight models are really open or built on stolen IP.
Moonshot AI seeks $50 billion valuation in pre-IPO round
Moonshot AI, maker of the Kimi model family, is reportedly seeking a $50B valuation ahead of a possible IPO this year.
Treasury Secretary says its 'unacceptable' that Chinese open source AI models may have been built on 'IP theft'
The US Treasury Secretary publicly called it unacceptable that Chinese open-source AI models may be built on stolen IP, escalating the dispute to a cabinet-level statement.
China's 'open' AI is a terrible business, and nothing like open-source software
A Business Insider analysis argues Chinese labs' "open" AI releases function as strategic loss-leaders rather than the community-governed open-source model familiar from software.
You are caught up for this edition