How DoorDash Built an AI Shopping Assistant That Doesn't Rely on the LLM Alone
Ask DoorDash pairs an LLM router with specialized agents, MCP-based tools, and persistent consumer memory instead of leaning on a single model for everything.
AI Daily Recap
11 articles · 4 categories
The finishable daily brief
What happened in AI — Jul 13, 2026
Monday, Jul 13, 2026
11 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Today's Show HN crop skews toward coding-agent tooling — control planes, browser vision, financial data feeds, and multi-agent attention tracking — plus an open P2P protocol for agent-to-agent coordination.
- DoorDash's Ask DoorDash assistant combines an LLM router, specialized agents, MCP tooling, and persistent consumer memory rather than one model doing everything.
- Google Cloud's k8s-aibom auto-generates AI bills-of-materials on GKE to catch unregistered "shadow AI" workloads that evade traditional scanners.
- AWS added a low-code UI on top of SageMaker's existing inference-recommendation API.
- Claude Fable 5 beat Hebbia's own finance-specific models on Hebbia's financial-diligence evals — their biggest accuracy gain yet.
Today's feed skews toward tooling for people who run coding agents day to day: control planes, browser-vision plugins, financial-data feeds, and attention trackers for juggling multiple agents at once, mostly from indie builders on Show HN.
The production-scale counterpart came from bigger platforms: DoorDash detailed a multi-agent shopping assistant with persistent memory, Google Cloud shipped AI bill-of-materials scanning for GKE, and Anthropic's Claude Fable 5 beat Hebbia's own finance-tuned models on Hebbia's internal evals.
Agent Runtimes, Orchestration & MCP 3 items
Production teams and open-source builders are converging on multi-agent architectures with explicit control layers — audited control planes, P2P coordination protocols, and memory-backed agent routing — instead of one do-everything model.
Show HN: PlanWright – A control plane for AI coding agents
MCP-based control plane lets you plan in Claude Desktop, implement in Codex, and review in a separate triage agent, with every decision logged for audit.
Node0 – A P2P protocol and SDK for autonomous AI agents
Open P2P protocol and SDK aim to let autonomous agents discover and transact with each other directly, without a central orchestrator.
Coding-Agent Developer Tools 4 items
A wave of small, focused tools is filling gaps in the coding-agent workflow: live financial data, remote browser control, visual feedback, and attention-routing across several concurrent agents.
Show HN: Finterm.ai Bloomberg terminal for Claude Code
New CLI gives coding agents direct access to stock prices, options data, and SEC filings, plus a filtered "Ticker Deep Research" mode, so agents can reason over live financial data without scraping.
Show HN: Call to Control AI Agents via the Web
Open-source ADE (Agentic Development Environment) client lets you drive Codex, Claude Code, or Open Code from a browser, so you can steer a coding agent away from your primary machine.
Show HN: Baton - Know which of your AI coding agents needs you
Baton tracks several running coding agents at once and surfaces which one is blocked waiting on you, addressing the babysitting problem of running multiple agents in parallel.
Show HN: Let your coding agent iterate by seeing the browser
peek-cli streams a live browser view into a coding agent's context so it can iterate on UI changes by seeing the rendered page, not just reading the DOM.
Infra, Security & Data Governance 3 items
Cloud providers and the local-first community are both tightening what counts as real control over AI systems — one making inference configuration easier to get right, the other making unregistered AI workloads and shallow ownership claims harder to hide behind.
Securing the AI supply chain on GKE: Introducing k8s-aibom for automated AI BOMs
Google Cloud's k8s-aibom automatically generates AI bills-of-materials on GKE to catch "shadow AI" workloads that developers deploy without registering them, which traditional scanners miss.
Launching UI for generative AI inference recommendations in Amazon SageMaker AI
AWS added a low-code UI on top of SageMaker's existing inference-recommendation API, so teams can get instance-type and configuration recommendations for a model without calling the API directly.
The Path to Sovereign Data: Challenges and Priorities in Local-First Computing
A panel on local-first computing argued that genuine data ownership requires structural independence and interoperability, not just account-level control, pushing back on how most "data ownership" claims are scoped today.
Models & Evals 1 item
A general-purpose frontier model outscored a domain specialist's own tuned models on that specialist's benchmark, a notable data point for teams deciding whether to fine-tune or just use a stronger base model.
Working at the frontier: How Hebbia builds AI for financial diligence that can't miss a detail | Claude by Anthropic
Anthropic's Claude Fable 5 outscored Hebbia's own finance-specific models on Hebbia's financial-diligence evaluations, the biggest accuracy jump Hebbia has recorded on that benchmark.
You are caught up for this edition