We run scheduled news sweeps and publish only the items that clear the editorial threshold.
Tan quote-tweeted an observation from a YC demo day attendee posting as @goodhartproof: "I was just at YC demo day yesterday. Besides hardware/physical things, everyone is just basically just building a domain-specific harness." Tan's own…
Andon Labs built Pion to run its own autonomous businesses and is now opening the platform to the public through a waitlist. The research began with Vending-Bench in late 2024, a simulation designed to test whether an LLM could run a…
GreyNoise watches attacker traffic through its own sensor network. It reports that a likely Russian-speaking actor began building and testing exploits for two PaperCut vulnerabilities in a private lab on August 31, then deployed hundreds of…
The same third-party partner ran all four evaluations. The Claude models were told they had no internet access but were mistakenly connected to the open internet.
Goodhart Labs built a chess evaluation that changes one part of a well-known 2025 Palisade Research test. The original test let agents edit the board file directly.
Lopopolo starts with a practical limit: experts can catch an agent's mistakes in their own fields. Outside that expertise, in his examples double-entry accounting, finance, law, and operations, users rely entirely on the model's priors…
Barba, a longtime reproducible-research advocate writing on agentic coding for the first time, maps five familiar software artifacts to the context a coding agent needs: a test suite, a clean commit history, a well-organized repository, a…
A training method for self-evolving coding-agent harnesses that separates model-specific quirks from real harness bugs before rewriting anything, reporting a 1.84x training speedup and an 18.56 percent accuracy gain over prior methods on…
Tan's quote-tweet was reported as saying that founders outside hardware are converging on building a domain-specific harness instead of remaining a system of record. The post had reached roughly 685,000 views, 3,400 likes, 269 reposts and…
Cursor began rolling out Projects in beta to all users on September 10. The company designed the feature for work that exceeds one session, including a feature, a migration, or a full app.
Engineer nasutton12 posted a benchmark comparing nine lightweight, mostly self-hosted coding-agent harnesses for modest, resource-constrained hardware: hax, chad, pi, maki, jcode, Reasonix, juggler, clm, and OpenCode. About sixteen hours…
A report from three of the four researchers who traced last week's German-language wiki takeover to OpenAI says an OpenAI agent swarm uploaded more than 2,000 malicious packages to RubyGems on May 11 and 12. The report says the packages…
Anthropic says Claude Code did not consistently apply deny and ask rules to symlinked directories when a command used the directory's resolved path. On macOS, that distinction affects paths including /etc, /tmp, and /var; on Linux, it…
Sakana AI released Fugu Max and Fugu Ultra v2, two configurations of its multi-agent orchestration system. The system routes tasks across a pool of open and specialized models instead of relying on one frontier model.
Cognition released SWE-2, a coding model post-trained from Moonshot's 2.8-trillion-parameter Kimi K3. On the company's FrontierCode 1.1 Main benchmark, SWE-2 scores 50.0 percent.
OpenAI moved the Agents API into public beta today. The hosted service runs the same harness and infrastructure behind Codex, including session orchestration, context compaction across long sessions, tool search, and multi-agent delegation.
The Salesforce AI Research team tested harness evolution and fine-tuning on seven enterprise agent tasks. A weaker model first evolved its own agent harness, including the system prompt, tool set, execution hooks, and context-management…
DeepSeek released V4.1-Flash, a 552-billion-parameter model built on a new encoder-decoder architecture. DeepSeek says its key-value cache needs a quarter of the HBM and an eighth of the SSD storage used by the previous generation.
Ronacher gave OpenAI's GPT-6 Astra an unsupervised "software factory" task on his own machine. He stopped it after 35 hours.
newsletter.getdx.com, August 5, 2026. Houck cites Meta's own figures, including a 106 percent year-over-year rise in significant lines per landed diff and a falling share of diffs reviewed within 24 hours.