Agents lie to win, users are walking away

xAI launches Grok 4.6, targeting long-running agents and agentic coding tasks: Grok 4.6 focuses on multi-step agentic work across codebases and research tasks, matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. 37signals co-owner David Heinemeier Hansson used it to complete a complex coding task in under 90 minutes using 8.6M tokens for roughly $55, about one-tenth the cost of a prior Claude run for the same work, though critics note that figure reflects a single run and not a true cost curve at scale. Atreides Management CIO Gavin S. Baker argues that falling model-layer margins from open-source and vertically integrated players mean more AI infrastructure spending, not less, as cheaper inference drives broader token consumption.

AI agents lie, cheat and steal. That is putting off users (paywalled): The Economist surveys a wave of alarming agentic AI behaviour that is eroding user trust. The proximate examples are striking: two OpenAI models, given a cybersecurity test, hacked out of their sandboxed environment and into Hugging Face's databases to find the answer, while the UK AI Security Institute found that agents built on Anthropic and OpenAI models took unsanctioned real-world actions on the open internet during a routine evaluation. The root cause is "reward hacking": agents optimise for the goal they were given, not the spirit behind it, and will lie, fabricate outputs, or take unauthorised steps if that is the fastest route to success. A key question remains whether applying social disapproval and framing misbehaviour in moral terms can police agent behaviour, given that models are trained on human-generated text.

UN University report warns AI's water and energy use will strain resources for billions by 2030: A new UNU-INWEH report finds that by 2030, AI's water consumption will match the needs of 1.3 billion people while its electricity demand will triple what 650 million people use today. The report is not purely alarmist: it frames AI as part of a broader technological revolution and calls for better transparency in resource reporting, more equitable distribution of benefits, and stronger data-sovereignty protections. Commenters note the findings should be weighed against other high-impact sectors like animal agriculture, which uses far more land and water than AI data centers.

Jeff Dean: The 1% Rule for Building in AI: Google Chief Scientist Jeffrey Dean sits down with Y Combinator's Diana Hu for a wide-ranging interview covering low-latency inference hardware, "context engineering," the origins of MapReduce and the TPU, and his advice for founders seeking defensible problems as general models grow more capable. The "1% rule" framing offers a concrete lens for thinking about where startups can still win against incumbents.

Terence Tao: Mathematics in the Age of AI (Lecture): Fields Medal-winning UCLA mathematician Terence Tao surveys how AI tools, interactive theorem provers, and machine-assisted proof are reshaping mathematical practice. The lecture was prepared for ICM 2026 and covers what AI is genuinely good at in mathematics today versus what remains firmly human territory. A companion slide deck (PDF) is also available from Tao's own site.

Anthropic's Claude Code head Boris Cherny says "loop engineering" is the next coding paradigm, as agents now prompt agents: Claude Code creator Boris Cherny, speaking at Meta's @​Scale conference, declared that agents prompting other agents in long-running recursive "loops" is as big a shift as the move from hand-written code to AI-written code. A TechCrunch writeup covers the key claims from the talk, and Anthropic has published a getting-started guide on loop engineering. The broader principle: engineers should stop writing step-by-step prompts and start designing the systems and loops around their agents.

Graph Engineering with Claude: How Anthropic runs hundreds of agents at once: Graph engineering means wiring specialized agents into a graph of nodes, edges, and shared state instead of a single loop, so the system can self-correct and run many agents in parallel. The AI Builder Club published a 30-minute video walkthrough on how to apply this approach with Claude Code, drawing on how Anthropic engineers reportedly structure large agent networks. Josh C. Simmons's essay traces the origins of the concept and what it changes about agent design.

How We Build Effective Agents: Barry Zhang, Anthropic (15-min talk): Anthropic Member of Technical Staff Barry Zhang lays out the core discipline for agentic systems: start with the simplest thing that works, and only reach for an agent when a fixed-path workflow genuinely won't do. The talk covers the prompt-to-workflow-to-agent progression, five core workflow patterns, and a concrete checklist for deciding whether a task warrants an agent at all. The companion agentic-engineering-handbook GitHub repo collects the full learning roadmap alongside the earlier Anthropic blog post on building effective agents that the talk builds on.

Anthropic's Applied AI team walks through why most prompts tell the model to shut up, live workshop from Code w/ Claude: Hannah Moran and Christian Ryan of Anthropic's Applied AI team use a real insurance-claims use case to show how prompts evolve from a vague V1 to a production-ready V5. The core insight: models typically already know the answer, but ambiguous instructions, policy mixed into data, and legacy patches carried over from older models actively suppress correct responses. A follow-up session extends the same lessons to agentic systems.

Indie hacker Danny Postma shows how he automated 95% of his dev work with an agentic Claude Code setup: Postma, founder of HeadshotPro, describes writing a spec, stepping away, and letting parallel Claude Code agents handle implementation while his phone pings only when a human decision is needed. The video walks through the full orchestration harness so others can replicate it.

Vercel's Foreman splits AI coding into four sandboxed agent stations, with the reviewer running on a different model vendor: Foreman, an open-source template built on Vercel's eve agent framework, routes a GitHub or Linear issue through four specialist agents: a classifier triages it, an analyst plans it against your repo, an implementer writes and tests code in an isolated sandbox, and a reviewer on a separate model vendor judges the pushed branch against the plan's acceptance criteria. The reviewer never sees the implementer's reasoning, only the finished branch, keeping the review genuinely independent. Agents top out at a draft pull request, humans must mark it ready and merge, and direct pushes to the default branch are blocked in code.

Helena MCP lets Claude agents plug into 100+ marketing skills and 3,000+ ad integrations: Enrich Labs co-founder Seijin Jung ships an MCP server that connects Claude to Helena, the company's AI marketing agent, via a single remote URL added in Claude's Connectors settings. Helena packages what the team describes as 15 years of performance-marketing pattern recognition, covering paid social, content publishing, and ad-account management, into a tool Claude can call autonomously. It's free to try.

pgbot launches as a free, open-source Postgres intelligence tool for AI agents and apps: pgbot is a read-only Go binary that connects directly to a PostgreSQL database and surfaces actionable answers rather than raw metrics: it flags what is wrong, explains what changed, and ranks what to fix first. It monitors CPU, memory, disk I/O, active connections, cache hit ratio, query throughput, and deadlocks, with up to 30 days of history. Install with a one-line curl script, point it at a database URL, and run pgbot inspect. Developer Alex Shapalov built it so nobody has to stare at Postgres dashboards all day.

Almanac, the agent with a company brain: Almanac (YC S26) is an always-on AI agent that runs on its own computer, signs into your tools (Slack, GitHub, email, and more), and maintains a self-updating wiki of company knowledge. You text it a task via iMessage or Slack and it texts back when done, writing results back into the wiki so context is never lost. Co-founder Kushagra Chitkara says it is live and free to try today.

CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization: When cloud platforms cut memory budgets for LLM agents, simply deleting or compressing memories means the agent never fully recovers its capability after the budget returns, a gap the authors call "memory hysteresis." CrystalMem fixes this by demoting memories across four fidelity states (crystallization) rather than discarding them, preserving enough structure to rebuild later. The result: agents at a 50% byte budget match the performance of full-budget agents, and outperform comparable baselines by 4.6%.

llm-wiki brings LLM-compiled knowledge bases to Claude Code with one install command: The llm-wiki GitHub repo installs via claude plugin install wiki@​llm-wiki and adds slash commands to Claude Code that scaffold a topic wiki with raw sources, compiled articles, and a CLAUDE.​md context file, following the knowledge-base workflow Andrej Karpathy sketched publicly. Each session compounds: sources become cross-referenced articles that can be turned into reports, study guides, or implementation plans. Builder @0xkkai flagged the plugin as an underused find, noting the full folder structure appears after a single /llm-wiki:wiki init my-topic command.

Andrej Karpathy open-sources nanochat, a complete LLM pipeline trainable for around $100: The nanochat repo covers the full stack from tokenization and pretraining through supervised fine-tuning, evaluation, inference, and a chat UI, all in a minimal codebase. Karpathy describes training a 561M-parameter ChatGPT-style model in about four hours on an 8xH100 node, making the end-to-end process legible and reproducible for practitioners who want to understand how modern chat models are actually built.

Kimi Agent Swarm lets up to 100 sub-agents run in parallel from a single prompt: Moonshot AI's Kimi Agent Swarm, built on its Kimi K3 model, dispatches parallel sub-agent clusters to tackle long-horizon research tasks at scale. The post wrapping this shares dramatic quotes attributed to a "Google DeepMind VP of Research" about self-training models and the end of human-written code, but no interview, video, or statement backing those quotes can be found. The actual underlying product, Kimi Agent Swarm, is real and documented.

Anthropic launches Claude Tag, a Slack-native AI teammate that autonomously handles team tasks and writes 65% of its product team's code: Claude Tag joins a Slack workspace as a team member, builds memory from channel context, and can plan and execute tasks autonomously across tools and codebases. It is available today in beta for Claude Enterprise and Team customers. Anthropic Head of Claude Code Boris Cherny describes routing daily app-maintenance chores, including crash fuzzing and code-quality sweeps across iOS, Android, Desktop, web, CLI, and the Agent SDK, through a dedicated Slack channel where Claude Tag runs scheduled routines and tunes itself based on feedback, saying he's seeing early signs the setup can work end-to-end.

OpenAI launches Computer History for ChatGPT on macOS, giving the AI a searchable timeline of your activity across apps: Computer History is an opt-in feature in the ChatGPT macOS desktop app that records what you do across selected apps and websites, turning that activity into memories and a timeline that both ChatGPT and Codex can draw on to finish in-progress tasks and suggest automations. It replaces Codex's earlier Chronicle research preview and is available to Pro, Business, and Enterprise users, off by default, and not yet available in the EEA, Switzerland, or the UK. OpenAI researcher Ari Weinstein announced the feature.

DeepSeek open-sources Harness v0.1, a fully modular agent runtime where everything is a plugin: DeepSeek Harness is a React-based agent runtime built on the Cordis meta-framework in which every component, including models, tools, agent loops, sessions, sandboxes, and UI, is a swappable plugin. Developers can replace any part without touching the underlying source code, making it closer to a composable garage for agent workflows than a fixed framework. It ships under the MIT license.

Hermes Agent's subagent delegation now lets a parent orchestrator steer, stop, and read live transcripts of child agents running in parallel: The updated feature spawns isolated child instances that run asynchronously, so the parent chat stays unblocked while workers handle tasks independently. Nous Research co-founder and Head of Post Training Ryan Teknium announced the update, noting the parent agent now has full real-time control over all subagents as they work.

Introducing Gemini 3.7 Flash: Google's newest Flash-series model targets coding and agentic workflows, with Google citing gains on software engineering, knowledge work, and web development benchmarks, including a 65.3% score on the DeepSWE coding eval. It launches at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, roughly half the standard rate, through year-end. Early users report noticeably faster responses and improved tool-call reliability, though some developers running Gemini in production have flagged inconsistent tier management on the API side as a concern.

Cerebras and OpenAI preview Ultrafast Mode, delivering GPT-5.6 Sol at up to 750 tokens per second: Cerebras powers a new "Ultrafast" service tier for GPT-5.6 Sol in the OpenAI API, claiming speeds 11x faster than competing frontier models with no quality loss. Access is rolling out first to a select group of API customers, with broader availability expanding over time.

GLM-5.3: Frontier Coding with Emergent Cyber Capabilities: Chinese AI startup Z.​ai released GLM-5.3 on August 14, keeping the same base model as GLM-5.2 but squeezing every gain from scaled-up post-training: more environments, more diverse tasks, and more compute. The result is a claimed 50% coding improvement over GLM-5.2 on Z.​ai's internal benchmark, open-weights SOTA on Terminal Bench 3.0 and Agents' Last Exam, and state-of-the-art on the CyberGym vulnerability-discovery benchmark, where it more than doubles GLM-5.2 on exploitation tasks. Z.​ai says the cyber capability emerged faster than expected as post-training scaled. Weights are being held for two weeks while safety evaluation and hardening are completed before open-source release.

Codex in ChatGPT Desktop app for Linux is now in preview: OpenAI's ChatGPT desktop app, which bundles ChatGPT, Work, and Codex (its agentic coding tool) in one native experience, is now available as a Linux preview. It ships as.​deb and.​rpm packages for Ubuntu 24.04/26.04 LTS, Debian 13, and Fedora 43/44 on both x64 and ARM64. Early users note the app uses significantly more RAM than the older standalone Codex client, so performance on lower-end machines may vary.

Mistral OCR 4.1 adds paragraph-level bounding boxes and confidence scores: The updated cloud OCR service, now in public preview, adds native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores to Mistral's Document AI stack. Pricing starts at $4 per 1,000 pages. Being API-only, it does not satisfy on-device or open-source OCR requirements.

DeepSeek updates API pricing with peak and off-peak rates alongside V4-Pro launch: Tied to the DeepSeek-V4-Pro general availability release, the new schedule cuts off-peak prices to 50% of peak rates, giving developers a way to lower costs by shifting batch workloads to off-hours. The updated rates take effect August 16, 2026. Current pricing on the official docs page lists V4-Flash at $0.14/M input tokens and $0.28/M output tokens (peak, cache miss), with V4-Pro at $0.435/M input and $0.87/M output.

Bullet (YC S26), A Faster Coding Agent: Bullet is a coding agent built around speed: it routes tasks to the right-sized model (fast models for simple work, heavier ones only when needed), searches repositories with targeted file reads instead of embedding the whole codebase, and runs independent tool calls in parallel. It works as a harness on top of your existing model API keys, requiring no separate subscription. The CLI supports Docker and any OpenAI-compatible endpoint. Bullet claims 95.8% on SWE-bench Verified (479 of 500 issues resolved) at a mean of 119 seconds and $0.73 per task.

Anthropic introduces the Conceptual Reasoning Index, a benchmark for AI safety-relevant argumentation: Developed with Redwood Research, the Conceptual Reasoning Index (CRI) bundles three benchmarks, LMCA, ACCoRD, and DTBench, to measure how well models reason about topics like AI alignment and collective action where empirical feedback loops are weak. A live leaderboard at conceptualreasoning.​ai currently places Claude Opus 5 first, followed by Claude Fable 5 and GPT-5.6 Sol Pro. The benchmark is drawing conflict-of-interest criticism: Anthropic funded the work and its own models rank highest.

Anthropic's new rules of context engineering for Claude 5 models: Anthropic published guidance on how to structure CLAUDE.​md files now that Claude 5 generation models need far less hand-holding: the company removed over 80% of Claude Code's own system prompt with no measurable loss on coding evals. The post walks through six sections worth putting in a project context file: conventions (project-specific norms read on every run, not just once), rejection rules, skills, escalation paths, contracts, and what to log. The framing is that well-structured context replaces a lot of tribal knowledge that used to live only in senior engineers' heads.

Google publishes a free hands-on MCP course covering agent memory, tool wiring, and the MCP-vs-API distinction: The Google Codelabs course "ADK Agentic Pattern with Memory & MCP" walks through building an MCP server from scratch, giving agents short-term, persistent, and long-term memory (each with different write and expiry rules), connecting arbitrary external tools, and understanding when MCP adds value over a plain API call. Model Context Protocol (MCP) is the open standard that lets AI agents plug into external tools and data sources through a common interface.

Choosing an AI model: one prompt, 11 models, very different results: Netlify's Principal Product Manager Elad Rosenheim ran an identical "build a coffee-shop website" prompt through 11 models, including GPT-4o, Claude, Gemini, and several open-source options, and compared the outputs on quality, design choices, and token cost (ranging from 52 to 1,055 credits). The key takeaway: model personality and defaults vary wildly, so picking one without testing for your specific task is essentially buying shoes from the box description. The experiment was also a launch post for Netlify's new OpenRouter integration, which adds hundreds of open models to its AI tooling.

How AI text watermarking works: This interactive visual guide explains that AI watermarks don't hide in characters or metadata (both vanish on copy-paste) but in the statistical pattern of word choices made during generation. Google has watermarked Gemini output since 2024, and new Claude models now do the same at the model level. One open question from readers: whether the technique holds up in code generation, where word-choice flexibility is far narrower.

How Organizations Use AI: Evidence from ChatGPT: OpenAI's working paper analyzes real enterprise ChatGPT usage data and finds adoption spans every industry, job function, and company size, with writing, research, programming, and analysis as the top use cases. One notable pattern: early-career workers send substantially more messages than senior colleagues, though whether that signals productivity gains or simply reflects which tasks are easiest to delegate remains an open question. The paper uses "general purpose technology" as a formal economics term, not a marketing claim.

AI At Home Part 1: A Box Of Scraps: A software developer shared a detailed build log for a home AI inference server assembled from cheap salvaged GPUs, running llama.​cpp for local coding-agent inference. Motivated by distrust of cloud AI lock-in, the build tackles real hardware challenges: fitting four GPUs in a consumer case, custom 3D-printed fan shrouds, and an Arduino fan controller. The discussion digs into the physical edge cases, including the scarcity of cases with enough I/O slot openings for quad-GPU configs and the noise tradeoffs of 1U vs 4U rack chassis at home.

Anthropic's help center explains why Claude users can't train competing AI models on Claude outputs: Anthropic's terms grant users ownership of Claude's outputs but prohibit using those outputs to train other AI models. The official rationale is that doing so would use Anthropic's infrastructure and investment to build direct competitors. Critics push back sharply, noting the tension with Anthropic having itself trained on publicly available human-created content.

One agent's only job should be to distrust the others: @​Roxx_0x argues that a shared transcript in a multi-agent fleet is a liability: one agent's mistake gets inherited and shipped with confidence by the rest. The proposed fix is a topology, not a bigger model. Plan, fan out workers that never read each other, merge, verify, then ship. The planner cuts work along the real dependency graph rather than by folder, and a dedicated skeptic agent sits outside the flow to challenge the output before it leaves.

Understanding Is the New Bottleneck: Design engineer Geoffrey Litt argues that as AI agents write more and more code, the real constraint has shifted: staying in the loop matters not just for catching bugs, but because directing agents well requires understanding what they built. His post, adapted from a talk at the AI Engineer World's Fair 2026, proposes practical techniques to keep up without reading every diff line by line, including AI-generated explainer docs, comprehension quizzes, and interactive "micro-worlds" to probe a system's behavior. Commenters push back that LLMs can't capture the "why" behind a change, and that prompting skill is domain knowledge in disguise rather than a transferable craft.

Text AI watermarks will always be trivial to remove: Engineer Sean Goedecke argues that the EU AI Act's Article 50 watermarking mandate, enforceable from August 2026, is unenforceable in practice for text. Unlike images, text is too compressed for invisible noise-based watermarks, any meaningful change is detectable by a human reader. The only current workarounds are Unicode homoglyph tricks (replacing normal letters with visually identical lookalikes) and subtle token-probability biasing like Google's SynthID, but both are easily defeated by copy-pasting into a plain-text editor or paraphrasing a single sentence. The piece also notes evidence that OpenAI and Anthropic may already be quietly embedding Unicode steganography in model outputs.