Show HN: Running over 80M tokens in one agent session with no compaction
A single agent session ran all 89 sequential Terminal-Bench 2.0 tasks — over 80M tokens — with no compaction and no measurable accuracy loss versus a fresh session per task.
AI Daily Recap
15 articles · 4 categories
The finishable daily brief
What happened in AI — Jul 14, 2026
Tuesday, Jul 14, 2026
15 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Google and partners published an open Agentic Resource Discovery spec for finding and verifying AI tools and agents.
- An agent ran all 89 Terminal-Bench 2.0 tasks in one 80M-token session with no compaction and no accuracy loss versus fresh sessions.
- LangSmith now traces coding agents across Claude Code, Codex, Cursor, and Copilot — tool calls, costs, and retries in one view.
- New open-source monitor "Agentmetry" catches coding agents reading ~/.ssh and phoning home.
- Codex usage grew more than 10x in six months to 7M users, adding 1M in a single day.
- Nvidia argues performance-per-watt, not chip count, now caps how much an AI factory can produce.
Two new open standards moved to formalize agent interoperability: Google's Agentic Resource Discovery spec for finding and verifying AI tools and agents, and a proposed cross-model operating contract for coding-agent tool calls. Google also shipped a Genkit Agents API with human-in-the-loop support, and one team ran an 80M-token single agent session with no compaction and no accuracy loss.
Agent trust tooling advanced too: LangSmith now traces coding agents end to end, a new monitor catches agents exfiltrating SSH keys, and AWS extended agentic QA into CI pipelines. Codex usage also passed 7M users as Anthropic published large-codebase and AI-native-org guidance.
Agent Orchestration & Standards 4 items
Two open specs pushed toward standardized agent interoperability — a discovery protocol for tools and agents, and a proposed cross-model contract for coding-agent tool calls — while separate work tested how far a single agent context can stretch.
Google's Genkit Ships Agents API with Detached Turns and Human-in-the-Loop for TypeScript and Go
Genkit's new Agents API packages message history, tool loops, streaming, and state persistence behind one chat() call, with detached turns for mid-flow human approval.
Google and Industry Partners Announce Agentic Resource Discovery Specification for AI Agents
Google and partners published Agentic Resource Discovery (ARD), an open standard for publishing, discovering, and verifying AI tools, APIs, and agents.
One Contract, Every Model: An Operating Standard for AI Coding Agents
A proposed operating-standard harness aims to give every model the same contract for coding-agent tool calls, cutting per-model glue code.
Evals, Observability & Agent Security 4 items
New tooling targets agent trust: LangSmith traces multi-framework coding agents end to end, a monitor watches for credential exfiltration, and AWS pushed agentic QA further into CI pipelines, alongside a new team-level agent-maturity benchmark.
How to Debug Coding Agents with LangSmith Traces
LangSmith now traces coding agents across Claude Code, Codex, Cursor, and Copilot, exposing tool calls, subagents, errors, costs, and retries in one view.
Agentmetry, catch your AI coding agent reading ~/.ssh and phoning home
Agentmetry is a new open-source monitor built to catch coding agents reading SSH keys or exfiltrating data over the network.
Accelerating software delivery with agentic QA automation using Amazon Nova Act – Part 2
AWS extended its Nova Act QA Studio with batch regression test suites and pipeline integration for parallelized, agent-run test execution.
Show HN: Benchmark your eng team's AI agent maturity in 5 minutes
A free 5-minute benchmark grades an engineering team's AI-agent maturity, built from data collected across hundreds of conversations with engineering leaders.
AI Infrastructure & Efficiency 2 items
Infrastructure coverage turned to constraints: Google Cloud detailed serving Claude at production scale across regions, while Nvidia argued power budget, not chip count, now caps AI factory output.
Claude at scale on Google Cloud: Frontier AI, built for enterprise production
Google Cloud detailed running Claude in production at scale — managing accelerators, holding latency steady across regions, and keeping regulated data in-region for long-context requests.
Why Performance per Watt Is the Ultimate Metric for AI Infrastructure Efficiency
Nvidia argues performance-per-watt, not raw throughput, is the metric that determines how many tokens a power-constrained AI factory can generate — and its revenue ceiling.
Coding Agents, Platforms & Adoption 5 items
Coding-agent competition sharpened as Codex usage passed 7M users while Anthropic published large-codebase and AI-native-org practices, and Oracle and OpenAI both pushed enterprise guidance for adopting agentic tooling.
How Claude Code Works in Large Codebases: Best Practices and Where to Start
Anthropic published best practices for running Claude Code effectively in large, unfamiliar codebases.
Running an AI-Native Engineering Org
Anthropic outlined how an AI-native engineering org restructures workflows and roles around coding agents rather than bolting agents onto existing process.
[AINews] Codex usage up >10x in 6 months to 7M users, +1M in the past ~day; did Codex overtake Claude Code??
OpenAI's Codex usage grew more than 10x in six months to 7M users, adding 1M in a single day — a scale signal in the coding-agent race with Claude Code.
Oracle opens Fusion Agentic Applications to pro-code developers and coding agents
Oracle opened its Fusion Agentic Applications to pro-code developers and coding agents, extending its no-code agent builder toward a programmable API surface.
How to manage AI investments in the agentic era
OpenAI's framework for managing AI investment centers on measuring useful work per dollar rather than raw usage, then scaling only the workflows that clear that bar.
You are caught up for this edition