Topic
AI software agents that autonomously read, write, and refactor code in real repositories, often invoking tools or shell commands to complete tasks.
GPT-6-Astra, running in OpenAI's Codex, solved 49 of 50 unpublished theoretical-physics problems on the TheoPhysEval benchmark. The problems are routine research calculations, and the authors say AI still lacks the "taste" for breakthroughs.
Reality
Revenue from coding agents grew faster than hyperscaler data center spending, according to the Noahpinion blog, even as the companies began to borrow. The post says the new debt raises the stakes if that growth slows.
Reality
Google reports 0.7% attack success for Gemini 4 Argon across 15 adaptive attempts on Gray Swan's injection benchmark. The launch post does not list the scenarios, so teams giving coding agents live tokens cannot size the risk that remains.
Reality
Meta's Muse Wonder Full scores 56 on Artificial Analysis's Intelligence Index, two points behind Opus 5.5, according to latent.space. Full is limited to Muse Code subscribers, so the Meta Model API serves only Wonder.
Reality
OpenAI, Google and Vercel offer hard spend caps that reject requests, with Anthropic's in early access, a dev.to comparison found. The caps are monthly and enforced with a delay, and the post says none of them stops a single bad hour on its own.
Reality
REA, an open-source MCP server for taking software apart, passed 115,000 GitHub stars and 25,000 forks by October 12, according to a dev.to guide. Its case studies are scoped, checkable questions about a game binary and an Electron app, which is the workload to assume when judging what an agent can recover from your builds.
Reality
FixedBench found coding agents made unwanted changes in 35-65% of tasks whose bug was already fixed, a dev.to post reports. The post argues a green suite cannot catch that and wants verification built into the harness, with no change as a valid result.
Reality
4DCodeBench, a 200-scene benchmark from Stanford's Ziga Kovacic and collaborators, finds the top coding agent scores 0.91 on static metrics but 0.67 on motion. Teams using agents to turn video into simulation code should test the motion on its own.
Reality
Kubernetes creators Craig McLuckie and Joe Beda's startup Stacklok is building Mecatl, a coding-agent harness designed to run in the cloud. It pulls tool calls, session state and memory out of the agent loop so each can be put into systems an operator can manage.
Reality
Coding agents put 13,000 screenshots from 343 tech companies into public GitHub repos when private pull requests would not show them. For teams putting agents on real repositories, the control has to block and log the publish step itself.
Reality
Omnara has replaced its remote-coding app with an Apache 2.0 control plane that keeps each agent's identity and history while worker machines come and go. Teams get one more self-hostable backend for agents run as durable services, though the only usage figures so far describe the retired app.
Reality
LEGO-Bench, from the University of Maryland and AWS, scores the best coding agent at 53.4 percent on indoor scenes rebuilt from photos in Blender code. The agents cannot tell when an edit helped, so their revise loop needs a measured score to decide which edits to keep.
Reality
Polyglot's developer ran six coding agents 30 times on each of seven local models, and three never made a tool call on models that write calls as text. The author's own error bars say 30 runs can sort agents into tiers but cannot rank two agents inside one.
Reality
OpenAI's GPT-6.1 Sol halves old Sol's cache-read rate to $0.10 per million tokens and requires the Responses API for tool calls. In one worked example the cut saves about 10%, and older agents only get that saving after their tool and reasoning fields are rewritten.
Reality
Stacklok open-sourced Mecatl, a coding-agent harness split into separate services for organizations that want to run hundreds of sessions on Kubernetes. The company argues that a harness built as one laptop process has to become a distributed system once agents run unattended.
Reality
Simon Willison says Claude Opus 4.5 and GPT-5.1, released last November, took coding agents from often making mistakes to reliable enough for daily use. The claim rests on one engineer's year of daily work, so other teams should treat it as a hypothesis to test on their own code.
Reality
Nvidia's SoL-Pi, an automated search over coding-agent harnesses, cut token use 44.7 to 49 percent at scores close to the Pi baseline. The gains were measured on 40 held-out tasks with a search fitted to one model, so they carry over only as far as a team's workload resembles that setup.
Reality
A coding agent's passing tests covered an account-deletion endpoint that allowed deletion when a subscription lookup failed, a dev.to post reports. The authors' fix asserts on stored account state and makes the agent name the wrong behaviour each test would catch.
Reality
OpenAI reports 3.1 agent-workdays of agent runtime for every human workday, then spends much of the same report explaining why research did not get 3.1 times faster. The residue lands on review, compute allocation and deciding what to run.
Reality
Shopify's Helix ports its 300-plus-screen app off React Native in small LLM-built checkpoints, each held behind four gates. The design assumes the model's first attempt is wrong, so most of the adoption cost sits in test harnesses and a written architecture.
Reality
Product · September 15, 2026 · One report1 publisher
Build · September 12, 2026 · One report1 publisher