Tech News — 2026-09-11

Introducing SWE-2: Pushing the Pareto Frontier

Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra

Cognition released SWE-2, a coding model achieving 50.0% on FrontierCode 1.1 Main within one point of Fable 5.1 while costing 64% less. The model scales reinforcement learning to multi-trillion parameters using a per-effort cost penalty algorithm that flattens the entire cost-performance frontier, not just peak scores. Post-trained from Kimi K3's 2.8T-parameter backbone, SWE-2 adds 5-6 benchmark points while cutting usage costs 81% versus SWE-1.7.

Why it matters

For engineers building or deploying AI coding agents, SWE-2 demonstrates that cost-efficient frontier performance is achievable through RL scaling and Pareto-aware training, not just larger inference budgets.

AI/ML / infoq.com

OpenAI Releases GPT-6 Astra for Coding and Computer Use

OpenAI released GPT-6 Astra, a model focused on computer use, coding, professional workflows, science, and cybersecurity, scoring 72.6% on OSWorld 2.0 and 74.1% on DeepSWE v1.1. It introduces an experimental context mechanism in Codex that maintains searchable notes across up to one million tokens, and is the first OpenAI model classified at the critical cybersecurity capability level under the Preparedness Framework. The production version restricts advanced offensive tasks, with defensive capabilities planned through the Daybreak program.

AI/ML / infoq.com

Session Traces and Cost Controls Help Diagnose AI Agent Failures

StackGen principal engineer Sabith K Soopy details how Langfuse session traces and cost controls diagnose AI agent failures, including repeated wrong-tool calls that standard monitoring misses. The post recommends asynchronous batch exporters to avoid blocking agents during telemetry outages, hard iteration caps and pre-execution checks to block identical consecutive tool requests, and appending tool calls to searchable logs with PII redacted. For Post-incident review, it advises exporting bounded operational metrics to Prometheus while using OpenTelemetry's semantic conventions for consistent tracing, noting that granular session context belongs in traces or logs, not metric labels.

AI/ML / dev.to

AI-Generated Tests Can Make Coding Agents Worse. Here's How to Check Yours

AI-generated tests can degrade coding agent performance, as shown in a preprint where weaker tests reduced SWE-bench Verified repair success by 3.9 percentage points. Poor tests amplify edge-case bugs like Python's shared falsy handling of None vs []—both pass the falsy check but encode different requirements. Fixing regression tests to assert identity (e.g., `is None`) instead of truthiness prevents agents from learning the wrong behavior.

AI/ML / thenewstack.io

AWS open-sources Pizza Bot: email-style inbox for background AI agents

AWS open-sourced Pizza Bot, a desktop application that provides an email-style inbox for managing background AI agents. The tool shifts from chat-based agent interaction to an asynchronous, threaded model where scheduled agents run autonomously, surface notifications, and request human approvals without requiring the user to be present. Built on LangGraph (via DeepAgents), it persists agent state in SQLite, supports multiple model providers (Anthropic, Bedrock, Gemini, OpenAI, Ollama), and runs self-hosted on macOS, Windows, Linux with browser and terminal clients.

Cloud / thenewstack.io

How AWS Lambda logs every flow across thousands of microVMs per host with eBPF and Rust

AWS Lambda replaced its aging network flow capture system with a purpose-built pipeline using eBPF and Rust to track every packet across thousands of Firecracker microVMs per host, each with fleeting millisecond lifespans. The old iptables-based approach broke at Lambda's density due to linear rule explosion—over 100,000 rules per worker for 2,000 micro-VMs—creating unbounded per-packet overhead. The new system prioritizes complete attribution (every flow tied to the correct tenant microVM) and near-zero performance overhead, as any extra CPU or RAM at Lambda's scale erodes utilization and operating margin.

Cloud / infoq.com

Terraform AWS Provider Continues Rapid Expansion as AWS Infrastructure Becomes More Complex

Terraform AWS Provider v6.62.0, released August 2026, adds support for Amazon DSQL, ECS, ECR, SES, Pinpoint, and enhancements to Bedrock AgentCore, CloudFront, and Secrets Manager, reflecting the provider's evolution from basic provisioning to a high-level abstraction for AI, resilience, and security services. The provider's rapid growth also underscores governance challenges, as evidenced by the withdrawal of v6.57.0 earlier this year, requiring enterprise teams to treat provider versioning like application dependencies. Meanwhile, v6.58.0 expanded observability with Prometheus anomaly detectors and scraper logging, while Pulumi, AWS CDK, and OpenTofu remain competing alternatives with different trade-offs in feature velocity, programming models, and licensing.

Languages / infoq.com

tsgolint Reaches Stable v7, Bringing Go-Powered Type-Aware Linting to Oxlint

tsgolint v7, the Go-powered type-aware linting engine behind Oxlint, is now stable, achieving 12-18x speedup over ESLint with typescript-eslint across large codebases like microsoft/vscode and vuejs/core. It builds real TypeScript programs on typescript-go (TypeScript 7's official Go port), covering 59 of 61 type-aware rules, and outputs structured diagnostics to Oxlint's Rust-based file discovery and syntax rules. The release tracks TypeScript v7.0.2, adds per-rule debug timings, and exposes compiler errors via --type-check.

Open Source / thenewstack.io

Shopify spent years on React Native — then rebuilt everything in 12 weeks

Shopify abandoned React Native after years of investment, rebuilding its Shop consumer app natively in 12 weeks using AI agents guided by an internal orchestration system called Helix. The company found that LLMs fundamentally changed the cost calculus of maintaining separate iOS and Android codebases, as agents could reliably generate platform-specific code and enable cross-platform engineering. This 180-degree pivot was not framed as a mistake, but as a strategic adaptation to the capabilities of modern coding agents, which allowed the team to skip years of gradual migration.

General / nasutton.notion.site

Nine coding harnesses vs. your laptop

A developer benchmarked nine coding AI harnesses against a local Qwen 3.8 27B model on an M4 MacBook Pro, finding three systemic problems: oversized system prompts and tool schemas (2,008 tokens for pi vs 18,046 for Opencode), reduced effective context windows (Opencode leaves only 44% of 32K tokens), and excessive side requests that overload the local GPU. The test used auto-approve Exercism tasks with a shared llama.cpp server, revealing local-local design flaws that make many harnesses unresponsive compared to data-center usage.