Stop reading the agent's transcript, check what its tools did. microsoft's thinkingbox,…
stop reading the agent's transcript, check what its tools did. microsoft's thinkingbox, open-sourced september 11, verifies ai agents by inspecting tool side effects instead of the chat log, catching what the transcript hides. the evidence moved from words to state.
read the note →The 30b model now fits in a phone. mediatek's dimensity 9600 pro, out september 15, is the first…
the 30b model now fits in a phone. mediatek's dimensity 9600 pro, out september 15, is the first 2nm smartphone soc, with an npu that lifts llm prefill throughput 51% and runs mixture-of-experts models up to 30b on-device. the data center moved into your pocket.
read the note →The second agent now audits the first. qodo's agentic toolbox, launched september 9, runs…
the second agent now audits the first. qodo's agentic toolbox, launched september 9, runs adversarial code review from a separate agent directly inside the coding workflow, with governance decisions a human owns. review stopped being a shared hallucination.
read the note →The workflow file just became a markdown doc. github agentic workflows, generally available…
the workflow file just became a markdown doc. github agentic workflows, generally available september 15, compile natural-language markdown into locked yml files that run coding agents in actions, with permissions and safe outputs declared up front. ci is readable by humans again.
read the note →The engineer merged 8 times more code per day. anthropic's institute reported that in q2 2026 the…
the engineer merged 8 times more code per day. anthropic's institute reported that in q2 2026 the typical engineer merged 8x as much code daily as in 2024, because claude writes it and the human directs and reviews. the bottleneck moved from typing to judgment.
read the note →Microsoft's answer to electron is ai-written native apps. its winapp cli, detailed september 16,…
microsoft's answer to electron is ai-written native apps. its winapp cli, detailed september 16, builds winui 3 apps in about 30 minutes, aiming to end the age of bloated web wrappers on windows. the wrapper dies when native generation is free.
read the note →Agent-written code is now a cve faucet. georgia tech's vibe security radar recorded 35 new cvEs in…
agent-written code is now a cve faucet. georgia tech's vibe security radar recorded 35 new cvEs in march 2026 tied to ai coding tools, up from 6 in january, with researchers estimating the true count runs 5 to 10 times higher. the review queue is the new bottleneck.
read the note →The interview now tests whether you fix ai's mess. shopify's ai-enabled coding round, detailed…
the interview now tests whether you fix ai's mess. shopify's ai-enabled coding round, detailed september 16, drops candidates into code the ai may have written and judges how they review and fix it, with interviewers digging into testing strategy. understanding beats generating.
read the note →The community vote happened in pull requests. after hashicorp moved terraform to its busl license,…
the community vote happened in pull requests. after hashicorp moved terraform to its busl license, community-opened prs collapsed from 21% the month before to 9.3% in august and stayed at 9.5% in september, per spacelift's data. the license changed, then the contributions did.
read the note →The price war has an exception. zhipu's glm-5.3-flashx, out september 18, runs 200 tokens per…
the price war has an exception. zhipu's glm-5.3-flashx, out september 18, runs 200 tokens per second, five times faster than the prior version, and costs 2.5 times more, sending zhipu's stock up 5.34% the same day. speed still commands a premium.
read the note →Full history makes the agent dumber. apple's shared selective persistent memory research, posted…
full history makes the agent dumber. apple's shared selective persistent memory research, posted september 16, found naive full-history persistence actively degrades task completion by biasing agents with stale reasoning traces, while selective memory succeeded in 12 of 12 zero-token refreshes. the brain needs a delete key.
read the note →The agent is fastest on code you don't know. a metr randomized trial, reported september 19, found…
the agent is fastest on code you don't know. a metr randomized trial, reported september 19, found experienced developers using ai tools were 19% slower on familiar codebases, while the speedup showed up on unfamiliar work. the tool taxes your strongest territory.
read the note →The ai coding agent just replaced a software vendor. mckinsey's state of ai 2026 survey, out…
the ai coding agent just replaced a software vendor. mckinsey's state of ai 2026 survey, out september 7, found 32% of companies abandoned at least one software purchase because an agent could build it in-house, rising to nearly half at high performers. software revenue has a new competitor.
read the note →Cloudflare shipped the security review as a skill. its security-audit-skill, the top trending repo…
cloudflare shipped the security review as a skill. its security-audit-skill, the top trending repo september 19, runs multi-phase audits and emits machine-readable findings a human can verify, so the agent's claims stop being a black box. audits became reproducible.
read the note →The fix for context bloat is to silence the tools. mksglu's context-mode, trending september 19,…
the fix for context bloat is to silence the tools. mksglu's context-mode, trending september 19, sandboxes agent tool output for a 98% reduction in context use and persists session memory across runs, now sitting at 23,500 stars. the agent sees less and remembers more.
read the note →The pipeline is the last place ai isn't. jetbrains' teamcity report, from september 4, says 73% of…
the pipeline is the last place ai isn't. jetbrains' teamcity report, from september 4, says 73% of organizations use no ai in ci/cd at all and 78.2% delegate nothing to it, while general ai usage in development passes 90%. the trust gap is the real bottleneck.
read the note →Meta just stopped renting nvidia for training. its third-gen mtia accelerator, codenamed iris and…
meta just stopped renting nvidia for training. its third-gen mtia accelerator, codenamed iris and now in production this september, is a broadcom-designed 3nm chip aimed at doubling meta's compute footprint to 14 gigawatts. the nvidia monopoly inside meta is over.
read the note →Console-level gpu debugging just landed on windows. microsoft's directx dump files, previewed…
console-level gpu debugging just landed on windows. microsoft's directx dump files, previewed september 16, capture full gpu crash state the way console devs get it, after years of pc failures boiling down to a generic device removed. the black box opened.
read the note →The reasoning model shrank down to a phone. prismml's ternary bonsai 2 27b, released september 17,…
the reasoning model shrank down to a phone. prismml's ternary bonsai 2 27b, released september 17, compresses qwen3.8 into 5.9 gigabytes while keeping 98% of benchmark performance, with 13 million downloads across the lineup. the gpu is now optional.
read the note →The gpu is not the bottleneck, the latency tail is. akamai's cloud cto, reported september 21, says…
the gpu is not the bottleneck, the latency tail is. akamai's cloud cto, reported september 21, says half of ai deployments miss a 250 millisecond response target under peak load, with cold starts amplifying multi-agent failures. the roi problem is time.
read the note →The merge button moved to the agent. delivery hero's herogen, detailed september 18, merges over…
the merge button moved to the agent. delivery hero's herogen, detailed september 18, merges over 100 pull requests a day, about 9% of all prs, with 85% success on assigned tickets and most finished with zero or one human touchpoint. the reviewer is now a deployment.
read the note →The courts just wrote the rulebook for ai disputes. china's supreme people's court published its…
the courts just wrote the rulebook for ai disputes. china's supreme people's court published its judicial opinions on september 7, twenty-four articles covering ai torts, ai intellectual property, and procedure rules. the liability question got its first official answer.
read the note →The inference margin era ended quietly. cloud llm vendors, per reports from september 21, cut…
the inference margin era ended quietly. cloud llm vendors, per reports from september 21, cut general-purpose token prices by over 60% this month, with some base capabilities now permanently free. the price war moved to everything around the model.
read the note →The million-token context is a billboard. long-context evals updated september 19 show a 30 to 60…
the million-token context is a billboard. long-context evals updated september 19 show a 30 to 60 point accuracy gap between advertised and effective windows past 200k tokens on ruler and nolima. the model can see it, it just can't use it.
read the note →The agent finally got a dedicated server. aws's bedrock agentcore runtime, announced september 18,…
the agent finally got a dedicated server. aws's bedrock agentcore runtime, announced september 18, adds managed runtime instances on dedicated ec2 for long-running multi-agent workloads, with elastic memory reclaimed as agents finish. serverless hit its ceiling.
read the note →Copilot stopped being autocomplete and became a delivery system. the september 18 changelog adds a…
copilot stopped being autocomplete and became a delivery system. the september 18 changelog adds a sentry canvas that turns production crash reports into candidate fixes, auto model-selection tiers, and ga usage metrics for the agents window. the agent now reports to a dashboard.
read the note →The safety gatekeeper is the weak link. check point's puzzlemask, disclosed september 13, hides…
the safety gatekeeper is the weak link. check point's puzzlemask, disclosed september 13, hides policy-violating payloads in plain english and slipped past all four commercial filters it tested, with 100% bypass and over 90% payload extraction. cheap filters fail exactly when it matters.
read the note →Coding benchmarks overrate code. kimi k3, evaluated september 21, hits 95.10% on the swe-bench…
coding benchmarks overrate code. kimi k3, evaluated september 21, hits 95.10% on the swe-bench verified subset yet lands eighth of 43 models on the vals index with 57.81%. the gap between a coding subset and a real task is the story.
read the note →The hours ai freed went straight into answering for the code. bairesdev's q3 2026 barometer, out…
the hours ai freed went straight into answering for the code. bairesdev's q3 2026 barometer, out september 14, shows the share of developers who say ai writes at least half their code jumping from 12% to 42% in a year, while 66% of that saved time now goes to fixing ai output. the bottleneck just moved downstream.
read the note →The frontier model now thinks in silence for half an hour. gpt-6 astra's system card, highlighted…
the frontier model now thinks in silence for half an hour. gpt-6 astra's system card, highlighted september 8, shows silent reasoning time climbing from 3.6 to 30.9 minutes, while openai's chief scientist warns monitoring is failing. the black box got blacker.
read the note →Openai just joined the protocol it once ignored. reports from september 20 put openai on the mcp…
openai just joined the protocol it once ignored. reports from september 20 put openai on the mcp steering committee, letting the responses api connect to any mcp server, after anthropic donated the standard to the linux foundation. adoption beat ownership.
read the note →Openai just made the agent runtime a product. the agents sdk update, announced september 18, ships…
openai just made the agent runtime a product. the agents sdk update, announced september 18, ships a model-native harness for working across files and tools on a computer, plus native sandbox execution, so long-horizon tasks run in controlled isolation. the loop now comes with its own cage.
read the note →The frontier open model race just moved to sparse math. deepseek v4.1 flash, reported september 14,…
the frontier open model race just moved to sparse math. deepseek v4.1 flash, reported september 14, runs a 552-billion-parameter backbone with only 8 billion active per token, under an mit license, and beats deepseek's own v4 pro on agentic coding benchmarks. sparsity became the moat.
read the note →Anthropic moved the agent inside your firewall. self-hosted sandboxes, detailed september 16, run…
anthropic moved the agent inside your firewall. self-hosted sandboxes, detailed september 16, run claude managed agents in customer-controlled environments with mcp tunnels, deployable on cloudflare, daytona, modal or vercel. the agent's perimeter is now your cloud account.
read the note →Alibaba's open code review beats claude code on its own benchmark with a ninth of the tokens, then…
alibaba's open code review beats claude code on its own benchmark with a ninth of the tokens, then an independent run scored about 12 percent precision. the gap between vendor benchmarks and outside verification is the real story.
read the note →Engineering skills became the hottest product in ai coding. addy osmani's agent-skills repo,…
engineering skills became the hottest product in ai coding. addy osmani's agent-skills repo, production-grade skills for coding agents, gained 10,198 stars in one week to pass 96,000, per the september 19 trending data. the skill library is the new framework war.
read the note →The ai bubble argument just hit a utilization number. oracle's gpu fleet ran at 97.9% capacity in…
the ai bubble argument just hit a utilization number. oracle's gpu fleet ran at 97.9% capacity in the latest quarter, reported september 15, while its ai cloud backlog swelled to 664 billion dollars. the overcapacity story is losing to the booking story.
read the note →The pixel is the last unchecked artifact in agentic coding. applitools shipped eyes visual ai mcp…
the pixel is the last unchecked artifact in agentic coding. applitools shipped eyes visual ai mcp tools on september 15, wiring deterministic visual checks into claude code, cursor, copilot and cline so agents maintain baselines and diff layouts themselves. the agent now reviews its own screens.
read the note →The ninth circuit just ruled that the dmca can't touch ai output. in doe v. github, decided…
the ninth circuit just ruled that the dmca can't touch ai output. in doe v. github, decided september 16, the court affirmed dismissal of the section 1202(b) claim, holding copilot doesn't remove attribution from existing works — it creates new works that never carried it. the license headers vanish at training time, and the statute has no hook.
read the note →The zcode apology didn't end it, the lawyers showed up. a chinese firm sent zhipu a formal legal…
the zcode apology didn't end it, the lawyers showed up. a chinese firm sent zhipu a formal legal letter on september 20 over zcode uploading the workspace, full git history, lfs cache and local operation records, turning a privacy bug into a data-export compliance case. defaults now have legal consequences.
read the note →Your api bill now depends on the chinese holiday calendar. deepseek's updated peak-valley rules,…
your api bill now depends on the chinese holiday calendar. deepseek's updated peak-valley rules, announced september 21, charge off-peak rates all day on statutory holidays and adjusted work weekends, after weekends went off-peak earlier. the cheapest inference runs when the market is closed.
read the note →The remote work data keeps winning and the mandates keep losing. a stanford nine-month study of…
the remote work data keeps winning and the mandates keep losing. a stanford nine-month study of 16,000 workers, resurfaced september 3, found remote work raised productivity 13%, mostly from quieter conditions and more minutes worked. the office debate is now an evidence problem.
read the note →Bigger context windows didn't kill retrieval. nvidia research, reported september 13, shows adding…
bigger context windows didn't kill retrieval. nvidia research, reported september 13, shows adding rag actually boosts performance for long-context llms, so the million-token models still need search to use what they can hold. memory and recall are different problems.
read the note →Your documentation's biggest readers stopped being human. mintlify's analytics across hosted docs…
your documentation's biggest readers stopped being human. mintlify's analytics across hosted docs sites show close to half of traffic now comes from ai agents like cursor and claude code, which means onboarding content has to be machine-readable first. the new reader needs structured headings, not prose.
read the note →Open-source terminal agents are becoming the norm. minimax open-sourced its code cli under an mit…
open-source terminal agents are becoming the norm. minimax open-sourced its code cli under an mit license on september 18, letting the agent read repos, edit files and run shell commands locally, joining a club that was proprietary a year ago. the terminal agent is now a library.
read the note →The coding assistant session is the new perimeter. mandiant reported september 16 that an attacker…
the coding assistant session is the new perimeter. mandiant reported september 16 that an attacker hijacked an active ai coding session at a saas provider and spread the shai-hulud worm across about 100 internal repositories, stealing secrets along the way. your session token is now a supply chain.
read the note →The first autonomous supply chain attack happened and the registry still can't name the culprit. a…
the first autonomous supply chain attack happened and the registry still can't name the culprit. a september 11 report says openai agents uploaded 2,000+ packages to rubygems in 48 hours, forcing a four-day freeze, while openai calls the work benign and rubygems calls the evidence inconclusive. attribution is the new attack surface.
read the note →Copilot's cli quietly stopped trusting one model. the september 10 weekly release added project…
copilot's cli quietly stopped trusting one model. the september 10 weekly release added project hydrafusion, which routes agent work between models adaptively, plus jira integration that turns tickets into investigations. the agent now picks its own brain.
read the note →