daily recap

OpenAI's Agents API launch reshapes the agent-infrastructure stack

The day is defined by OpenAI's September 10 push into managed agents: a public-beta Agents API built on the Codex harness plus a full-duplex GPT-Live-1 voice model, which together pulled a wave of sandbox and orchestration partners into orbit. E2B and Temporal shipped execution backends for the Agents SDK, framing a contest over where long-running agents actually run. In parallel, a second theme centered on making agents trustworthy and reliable rather than just capable: verifiable coding runtimes (Atomic, hcode), regression-style agent evaluation (LangChain and Harbor with LangSmith), and a cross-network Know-Your-Agent identity framework from Visa, Mastercard, and Ant International. Cognition's SWE-2 model and PlanetScale's sharded-Postgres preview round out the day on the model and data-infrastructure fronts.

This daily brief links to the published stories that shaped the day's market picture.

Published Sep 11, 2026 Updated Sep 11, 2026

daily-roundup

Summary

Key themes

OpenAI moves into managed agents: the Agents API public beta exposes the Codex harness behind one call (session state, context compaction, crash recovery, subagent coordination), while GPT-Live-1 brings a full-duplex voice front end at $0.05/minute.
A sandbox-backend contest forms around the Agents SDK: E2B (as a launch partner) and Temporal both shipped isolated execution environments, competing on isolation, durable/replayable state, and parallel runs as orchestration shifts to managed APIs.
Reliability and verification over raw capability: Atomic's open-source verifiable runtime and hcode's local role-isolated IDE make the engineering process itself checkable, while LangChain and Harbor turn agent evaluation into parallel regression testing via LangSmith.
Agent trust becomes shared infrastructure: Visa, Mastercard, and Ant International agreed to align on a cross-network Know-Your-Agent framework to identify AI agents in agent-initiated commerce.
Model and data-infrastructure progress continues: Cognition's SWE-2 (post-trained from Kimi K3) claims near-frontier coding scores at ~64% lower cost, and PlanetScale previewed Neki, horizontally sharded Postgres behind a single connection string.

Notable items

OpenAI opens the Agents API public beta on the open-source Codex harness, with no added API fee beyond model tokens and paid tools.
OpenAI launches GPT-Live-1 in the API, a full-duplex voice model at $0.05/minute reporting a 30-point Full Duplex Bench gain over GPT-Realtime-2.1.
E2B becomes a launch-partner sandbox backend for the OpenAI Agents SDK, running SandboxAgent sessions with full filesystem, terminal, and network access (Python SDK first).
Temporal adds agentic sandboxes pairing durable execution with isolation, giving long-running agents crash recovery and replayable state.
Cognition releases SWE-2, post-trained from Moonshot Kimi K3, scoring 50.0% on FrontierCode 1.1 Main (near Anthropic Fable 5.1) at roughly 64% lower cost, shipping in Devin Desktop and CLI.
Visa, Mastercard, and Ant International align on a Know-Your-Agent interoperability framework bridging their respective agent protocols while each keeps its own verification.
Atomic ships an open-source verifiable coding-agent runtime with stages, checks, and approval gates across Codex, Claude, Copilot, and more.
LangChain and Harbor detail a unified LangSmith evaluation stack that isolates each agent trial and streams full traces for parallel, rerunnable regression testing.
PlanetScale opens a platform preview of Neki, horizontally sharded Postgres from the Vitess team, with routing, resharding, and online schema changes managed for you.
hcode launches a local desktop IDE (Tauri/Rust/React) that runs Claude, Codex, and Grok in slots under a disciplined SDLC with a deterministic merge gate and human sign-off.

Source coverage

Source rows used: 10

Daily brief

Top stories behind today's brief

These published stories shaped the day's roundup. The ranked list keeps the most important changes easy to scan.

01

Sakana AI launches Fugu Ultra v2 and Fugu Max, orchestrating a pool of models beh…

Sakana AIFugu Ultra v2agentsfeature launchmedium impact

On September 11, 2026, Sakana AI released Fugu Ultra v2 (capability-first) alongside Fugu Max (cost-first). Fugu is a multi-agent system delivered as a single OpenAI-compatible model that dynamically routes and orchestrates a pool of underlying models. Fugu Ultra v2 has a 1M-token context and leads five of eight benchmarks in Sakana's own evaluation, including 48.3 on Chartography versus Claude Opus 5 at 27.3.

2 sources

agentsai-coding-toolsfeature launch
02

Salesforce launches seven job-ready Agentforce agents for sales, service, IT, and…

SalesforceAgentforceagentsfeature launchhigh impact

On September 11, 2026, ahead of Dreamforce, Salesforce expanded Agentforce with seven pre-built, job-ready agents, Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin, spanning roles such as customer service and IT/HR support. Six are generally available immediately; the outbound sales agent Hunter is in pilot with general availability targeted for November 2026.

2 sources

agentsenterprise-controlsfeature launch
03

Vercel adds connector-management permissions to Vercel Connect…

VercelVercel Connectenterprise controlsfeature launchlow impact

On September 11, 2026, Vercel added access controls to Vercel Connect so Pro and Enterprise teams can restrict who is allowed to create and manage connectors, tightening governance over how projects connect to external services and data sources.

1 sources

enterprise-controlssecurityfeature launch
04

Atomic ships an open-source verifiable coding-agent runtime with stages, checks,…

AtomicAtomicai coding toolsopen source releasemedium impact

Around September 10, 2026, Atomic (by bastani-inc) released an open-source coding-agent runtime that lets teams define an engineering process in natural language as explicit execution graphs with stages, checks, and approval gates, rather than relying on prompt instructions. Runs can span hours to weeks as the runtime researches, implements, and verifies work across models such as Codex, Claude, and Copilot.

1 sources

ai-coding-toolsagentsenterprise-controls
05

Cognition releases SWE-2 coding model, post-trained from Kimi K3 and shipped in D…

CognitionSWE-2ai coding toolsopen source releasehigh impact

On September 10, 2026, Cognition released SWE-2, its most capable coding model, post-trained from Moonshot Kimi K3 using reinforcement learning that trains medium, high, and max effort levels in a single run. SWE-2 scores 50.0 percent on FrontierCode 1.1 Main, within about a point of Anthropic Fable 5.1, while Cognition says it costs roughly 64 percent less to run at that score. It is available in Devin Desktop and Devin CLI at launch.

2 sources

ai-coding-toolsagentsrepo-momentum
06

E2B becomes a launch-partner sandbox backend for the OpenAI Agents SDK

E2BE2B Sandboxesagentsintegrationmedium impact

Alongside the OpenAI Agents SDK and Agents API launch on September 10, 2026, E2B shipped a native integration that runs agent sessions inside isolated E2B sandboxes with full filesystem, terminal, and network access. Agents can keep temporary workspace state across steps, produce reviewable artifacts, and run many sandboxes in parallel; sandboxing is supported first in the Python Agents SDK.

2 sources

agentsenterprise-controlsai-coding-tools
07

LangChain and Harbor ship a unified stack for evaluating long-running agents…

LangChainLangSmithenterprise controlsintegrationmedium impact

Around September 10, 2026, LangChain and Harbor detailed a unified evaluation stack that runs each agent trial in an isolated LangSmith sandbox and streams results into LangSmith observability with full agent traces. Teams can run hundreds of trials in parallel and rerun evaluations when skills, models, or tool descriptions change, turning agent evaluation into regression testing.

2 sources

enterprise-controlsagentsintegration
08

OpenAI launches GPT-Live-1 in the API, a full-duplex voice model at $0.05 per min…

OpenAIGPT-Live-1agentsfeature launchmedium impact

On September 10, 2026, OpenAI launched GPT-Live-1 in the API, a full-duplex voice model that listens and speaks at the same time and delegates deeper reasoning and actions to paired models and tools. OpenAI priced the voice front end at $0.05 per minute and reported a 30 percentage point gain on Full Duplex Bench over GPT-Realtime-2.1, with large improvements in turn-taking latency.

2 sources

agentsai-coding-toolsfeature launch
09

OpenAI opens Agents API public beta, exposing the Codex harness behind one API ca…

OpenAIAgents APIagentsfeature launchhigh impact

On September 10, 2026, OpenAI launched its Agents API in public beta for all developers, built on the open-source Codex harness. The API handles session state, context compaction, crash recovery, and subagent coordination, so developers supply tools, instructions, and an execution environment while OpenAI manages the orchestration loop. There is no added API fee beyond model tokens and paid tools.

2 sources

agentsai-coding-toolsenterprise-controls
10

PlanetScale opens platform preview of Neki, horizontally sharded Postgres

PlanetScaleNekienterprise controlsplatform expansionmedium impact

On September 10, 2026, PlanetScale opened a platform preview of Neki, horizontally sharded Postgres built by the team behind Vitess. Every shard is real Postgres behind a single connection string, with PlanetScale handling routing, resharding, online schema changes, and version upgrades. Organization admins can opt in from the dashboard under Platform Preview.

2 sources

enterprise-controlsagentsplatform expansion

Citations

Related source citations

These primary sources come from the individual stories behind this daily brief.