Some company news and a quick ask before this week’s stories! Since 2019, we’ve helped developers transition into AI engineering and trained enterprises to build with AI. Over those seven years, enterprise demand moved past training; the ask now is to build and deploy production systems.
This week, we formalized that work as a dedicated effort: Towards AI Deployment.
We would really appreciate it if you could please like and share our co-founder’s post on LinkedIn and watch the video he has shared!
Towards AI now works as two connected sides: Learning converts software developers into AI engineers and forward-deployed engineers, and Deployment puts them to work delivering custom systems for private equity firms, their portfolio companies, funds, and banks. We believe successful AI deployment requires domain expertise, so we have narrowed our focus to a vertical where we have strong momentum and the domain expertise to complement our AI talent.
For readers of this newsletter, the most direct change is that the teaching improves. Every engagement the division delivers sharpens the curriculum: more failure cases, evaluation methods, and architecture patterns flowing back into our courses and these pages.
If you know of a company stuck between experimenting with AI tools and building a system people rely on, please [reach out](https://towardsai.com/), we’d love to look at building the workflow with you.
This was a huge week for model releases. In the span of two days, SpaceXAI launched Grok 4.5, OpenAI moved GPT-5.6 from restricted preview to general availability, and Meta launched Muse Spark 1.1. OpenAI also shipped GPT-Realtime-2.1 and a mini variant, improving interruption handling, noisy audio, and alphanumeric recognition while cutting p95 latency by at least 25%. Three new frontier competitors landed almost at once.
The week’s more consequential development was a price-performance reset. Closed models already led on raw intelligence; open weights led on cost per unit of it, and that second advantage slipped this week. GLM-5.2, the leading open-weight model on the Artificial Analysis Intelligence Index, scores 51 at a measured cost of about $0.37 per benchmark task. Grok 4.5 scores 54 at about $0.31, beating GLM on both score and cost. GPT-5.6 Luna and Muse Spark 1.1 tie GLM at 51 while costing about $0.21 and $0.26 per task. Sol and Terra score 59 and 55 at higher task cost. Closed models now match or beat the open-weight frontier on cost per unit of intelligence, which had been open weights’ clearest selling point.
Cost per task matters more than the price printed per million tokens. Grok 4.5 charges more per output token than GLM-5.2, but it used roughly 14,000 output tokens per benchmark task against GLM’s 43,000, so more intelligence per token beat a lower token price. Simon Willison’s identical SVG prompt cost 0.71 cents on Luna with reasoning off and 48.55 cents on Sol at maximum effort, a 68x spread from settings alone. In production, the number to watch is cost per accepted result; more on that below.
We covered the GPT-5.6 preview two weeks ago, including the Sol, Terra, and Luna product ladder, pricing, safety restrictions, Ultra mode, and METR’s difficult-to-interpret time-horizon result. The incremental news this week is stronger: the family is now generally available, independent tests are in, and they support OpenAI’s core capability and efficiency claims.
Sol is seriously competing with Claude Fable 5 at the frontier again. It scores 59 on the Artificial Analysis Intelligence Index against Fable’s 60, leads the Coding Agent Index, and performs particularly well on DeepSWE, Terminal-Bench, BrowseComp, and OSWorld. Fable stays ahead on SWE-Bench Pro, GDPval, Toolathlon, FrontierMath Tier 4, and broader professional-work comparisons. I would choose between them based on the job. Sol is exceptionally good at coding, computer use, presentations, and structured execution. In my own use, Fable still writes better, and Arena’s preliminary creative-writing leaderboard points in the same direction, at 1507 versus 1486 with overlapping uncertainty.
Luna may be the most useful release in the family. It ties the open-weight intelligence frontier, scores 75 on the Coding Agent Index through Codex, runs at more than 200 output tokens per second, and costs less per measured task than GLM-5.2. CodeRabbit provides a useful warning at the other end of the family: Sol passed 63.7% of more than 100 repository tasks without an execution error, yet its code-review precision was only 31.6%. Long-running execution is improving faster than verification.
Adoption moved almost as quickly as the models. OpenAI reported more than 5 million weekly active Codex users in early June. In the days after launching ChatGPT Work and combining Chat, Work, and Codex in one desktop app, OpenAI’s Codex lead reported 8 million active users across Codex and ChatGPT Work. The definitions differ, so this is a directional comparison. More than 1 million people were already using Codex outside software development before the Work launch, and the new interface gives that audience a much friendlier route into the same agent infrastructure. This may be the strongest signal in the release: the distribution layer is catching up with the capability layer.
Grok 4.5 closed the gap faster than I expected. It now sits in the same broad agentic-coding group as GPT-5.5 and Fable while costing much less, serving at around 80 tokens per second, scoring strongly on SWE Marathon and Terminal-Bench, and using far fewer tokens than several peers. Snorkel measured a 29% full-rubric pass rate across roughly 2,000 professional tasks, ahead of GPT-5.5 and Opus 4.8. Grok ran in its own Grok Build harness while competitors used a different agent, so treat that as a system comparison. It still trails Fable and GPT-5.5 on DeepSWE 1.1. Cursor has flagged a contaminated CursorBench result, and its 54% hallucination rate on AA-Omniscience means it needs oversight. I see it as a credible frontier model with excellent economics, one tier below the very best coding systems.
The Cursor connection clearly helped shape it. Cursor and SpaceXAI say they jointly trained Grok 4.5 on trillions of tokens of eligible Cursor data covering developers, codebases, tools, and agent interactions, with Privacy Mode sessions excluded. SpaceXAI also ran reinforcement learning on hundreds of thousands of multi-step technical tasks, with rollouts lasting many hours across tens of thousands of GB300 GPUs. There is no public breakdown of how much of the capability jump came from Cursor’s data versus the scale and design of the rest of the training recipe, but this is a serious product-data flywheel.
Muse Spark 1.1 made a similar jump, improving Meta’s Artificial Analysis score by eight points in three months, with a one-million-token context window and explicit training for both sides of a multi-agent system: a lead agent that plans and delegates, and a subagent that follows a narrow remit and escalates when needed. Meta’s own evaluation report is more balanced than the launch copy. Muse posts 88.1 on MCP Atlas and 80.8 on OSWorld-Verified for tool and computer use, but 53.3 on DeepSWE and 61.5 on SWE-Bench Pro, trailing the strongest GPT and Claude models on the longest cross-application workflows. Meta itself says current autonomy is insufficient for sustained automated AI research. What matters here is the rate of progress and the price. Grok 4.5 and Muse Spark 1.1 join GLM-5.2 in a cluster of releases that suddenly make long-running agentic coding much cheaper and more accessible.
The timing makes me wonder whether several labs recently gained access to the same new generation of long-horizon training environments from vendors such as Mercor or Surge, both of which publicly build repository-scale coding tasks, long tool-use trajectories, and reinforcement-learning environments for leading labs. Scaled in-house reinforcement learning, better harnesses, more compute, and stronger first-party product data are also contributing.
The next step change is multi-agent
The more I use GPT-5.6 Ultra and Fable, the more this looks like another step change in capability and in how easily we can consume huge amounts of compute. The first change was o1, which let a model spend far more computation on a single response. The next was Claude Code and Codex, where models such as Opus 4.5 and GPT-5.1 made multi-hour tasks practical with tools, persistent state, and automatic context compaction. Almost every month since last November, a new model or harness has stretched the length of work I can hand off. Now the unit of work is changing again: from one answer, to one long-running agent, to a managed team of agents.
GPT-5.6 Ultra coordinates four agents by default, and OpenAI published evaluation runs with 16. Its Responses API lets a root model decide when to delegate and then synthesize the work. Claude’s dynamic workflows can run tens to hundreds of parallel subagents when enabled, and Muse Spark 1.1 was explicitly trained to operate as both manager and worker. Before this, making many parallel agents behave like a competent team required a very strong developer: dividing the task, constructing prompts, isolating environments, preventing duplicate work, and merging results. These systems move much of that orchestration into the model and product. Ten to 100 agents is now a realistic configured scale, and Anthropic describes workflows with hundreds. However, fan-out at this scale still produces duplicated work, conflicting changes, security risks, and confident group mistakes.
I could now very easily spend $1,000 on a single prompt on either Claude or Codex, meaning one top-level objective that triggers hundreds or thousands of model calls across a team. Fifty Sol agents each reading 500,000 fresh input tokens and producing 50,000 output tokens cost about $200; several deep passes, retries, and tool calls, or a 100-agent configuration, push the same objective toward $1,000. Prompt caching softens this less than it appears, because workers on separate subtasks mostly read fresh context, cache writes carry a 1.25x surcharge, and reasoning and output tokens are never discounted. Anthropic’s 16-agent compiler project cost almost $20,000, so this is already more than a spreadsheet thought experiment.
Cost control is becoming part of prompt design: total budgets, per-agent limits, maximum fan-out, cheaper worker models, checkpoints before another round, and independent review before accepting the final synthesis. The next generation of power users will pair prompt-writing with disciplined compute allocation.
This week’s model pile-up is good news for token buyers. Sol and Fable now trade wins at the frontier, while Grok 4.5, Muse Spark 1.1, and increasingly capable open-weight models are creating credible competition on intelligence, speed, and cost per completed task. That pressure should keep lowering costs while increasing the amount of useful work each token can buy. The next constraint is training data. Coding agents improved because labs captured long repository trajectories: how developers explore unfamiliar systems, choose tools, recover from failed approaches, verify changes, and divide large goals across teams of agents. ChatGPT Work shows where this should expand next. We need equivalent multi-hour and multi-day task data across research, finance, legal, operations, marketing, sales, and other professional work, capturing the decisions, failures, corrections, and final outputs. Organizations should begin collecting these traces with consent, strong privacy controls, and rigorous evaluation. The labs that learn how expert work unfolds over hours or days will train the strongest general-purpose agents, and the growing competition among them should make those capabilities cheaper for everyone buying tokens.
*— *Louie Peters — Towards AI Co-founder and CEO
- Meta Releases Muse Spark 1.1 Meta Superintelligence Labs released Muse Spark 1.1 on July 9, a multimodal reasoning model designed for agentic tasks requiring planning and orchestration across external apps and services. The model accepts text, image, video, PDF, and audio inputs through a 1M-token context window with active compaction, meaning it remembers actions, retrieves information from earlier work, and filters out noise while preserving critical steps. Muse Spark 1.1 zero-shot generalizes to new native tools, MCP servers, and custom skills. It can run either as a primary agent delegating to parallel subagents or as a subagent itself. For computer use, it autonomously decides when to write a script for automation versus clicking through a GUI. On Meta-reported benchmarks, it leads on agentic orchestration: 88.1 on MCP Atlas, 54.7 on JobBench (Opus 4.8: 48.4), and 62.1 on Humanity’s Last Exam with tools (Opus 4.8: 57.9). It trails on coding: 61.5 on SWE-Bench Pro (Opus 4.8: 69.2) and 53.3 on DeepSWE 1.1 (GPT-5.5: 67.0). Alongside the model, Meta launched the Meta Model API in public preview for US developers, the first time Meta has offered a paid hosted model. API pricing is $1.25/$4.25 per million input/output tokens with $20 in free credits. The API supports both OpenAI and Anthropic wire formats. Muse Spark 1.1 is free in the Meta AI app in Thinking mode.
SpaceXAI released Grok 4.5, built for coding, agentic tasks, and knowledge work. The model was trained alongside Cursor on datasets spanning coding, science, engineering, and math, with reinforcement learning covering hundreds of thousands of tasks centered on multi-step software engineering. RL training runs asynchronously, with agentic rollouts that can last for many hours while learning continues across tens of thousands of NVIDIA GB300 GPUs. SpaceXAI highlights token efficiency as the primary advantage: Grok 4.5 resolves SWE-Bench Pro tasks with an average of 15,954 output tokens, roughly 4.2x fewer than Opus 4.8 (max), which uses 67,020 tokens. The model serves at 80 tokens per second. On benchmarks: 83.3% on Terminal-Bench 2.1, 64.7% on SWE-Bench Pro, 62.0% on DeepSWE 1.0, and 29.0% on SWE Marathon (first place). Fable 5 leads on SWE-Bench Pro (80.4%), DeepSWE 1.0 (66.1%), and Terminal-Bench 2.1 (84.3%). Beyond coding, Grok 4.5 handles office work through Grok Build, including multi-sheet Excel models with web research, PowerPoint diagrams using native shapes, and Word document drafting. Pricing is $2/$6 per million input/output tokens. Available in Grok Build, Cursor on all plans, and through the SpaceXAI API console, with free usage for a limited time in Grok Build and Cursor. Not yet available in the EU; expected mid-July.
-
OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2 OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini, two developer-facing speech-to-speech models for building voice agents through the Realtime API. The release reduces p95 latency by at least 25% across all Realtime voice models through improved caching. GPT-Realtime-2.1 updates its predecessor with improved alphanumeric recognition for items like order numbers, phone numbers, and confirmation codes, better silence and noise handling, and more reliable interruption behavior when a user speaks over the model. It adds configurable reasoning effort per request. GPT-Realtime-2.1-mini is a smaller reasoning model for faster, lower-cost voice interactions, shipping reasoning and tool-use capabilities at mini pricing for the first time. Pricing: GPT-Realtime-2.1 at $32/$64 per million audio input/output tokens; GPT-Realtime-2.1-mini at $10/$20. Both are available through the Realtime API via WebRTC, WebSocket, or SIP connections.
-
OpenAI Releases GPT-Live and GPT-Live-1 mini OpenAI launched GPT-Live, replacing Advanced Voice Mode as the default voice experience in ChatGPT. GPT-Live is a full-duplex audio language model that processes incoming speech and generates outgoing speech concurrently, rather than waiting for a user to finish before formulating a response. It interprets intonation and conversational intent, not just words. When a question exceeds GPT-Live’s native capabilities (complex math, multi-step research, tasks requiring tool use), it silently delegates to GPT-5.5 running in parallel without breaking the conversation. Users can select from three reasoning levels: Instant for quick replies, Medium for moderate depth, and High for thorough analysis. GPT-Live-1 is the default for Go, Plus, and Pro users; GPT-Live-1 mini is the default for Free users. On OpenAI’s benchmarks, GPT-Live-1 (High) scored 84.2% on GPQA (Advanced Voice Mode: 45.3%) and 75.2% on BrowseComp (Advanced Voice Mode: 0.7%). The launch includes nine remastered voices. Video and screen sharing are not supported at launch. API access is planned but not yet available; developers can sign up to be notified.
-
Prime Intellect Releases Verifiers v1 Prime Intellect released verifiers 0.2.0, previewing a rewritten core under the verifiers.v1 namespace. The redesign solves two structural problems: a monolithic environment design that coupled data, agent logic, and infrastructure, and a trace format in which storage grew quadratically with the number of turns (each turn stored a full copy of the prompt alongside the new completion). Verifiers v1 decomposes environments into three independently swappable layers: a taskset defining data, tools, and scoring; a harness defining how the task is solved (ReAct loop, Codex, Kimi Code, Terminus 2, Mini-SWE-Agent, or custom); and a runtime defining where execution happens (local subprocess, Docker, or cloud sandboxes including Prime Sandboxes and Modal). A directed acyclic graph message format replaces the old prompt-completion pairs, eliminating quadratic growth and enabling training on trajectories longer than the agent’s native context window through DAG branching. The library supports OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages wire formats via dialect adapters. Harbor is the first fully supported third-party taskset format, with alpha support for NeMo Gym and OpenEnv. The legacy v0 code path is frozen and will not be actively maintained.
-
Anthropic Discovers Global Workspace in Claude (J-Space) Anthropic published research identifying the J-space, a small collection of internal neural activation patterns in Claude that function as a shared mental workspace. Named after the Jacobian-based technique (J-lens) used to find it, the J-space holds concepts the model is thinking about but not saying. It was not designed or programmed; it emerged during training. When Claude reads code with a bug, “ERROR” appears in the J-space. When it reads a prompt injection, “injection” and “fake” appear. When solving multi-step math, intermediate answers appear in sequence. Experiments demonstrate four functional properties: Claude can report J-space contents (swapping “Soccer” for “Rugby” changes what it reports thinking about); modulate them on request (told to think about citrus fruits while copying an unrelated sentence, “orange” and “fruits” appear); reason with them (swapping “spider” for “ant” changes an answer about leg count from 8 to 6); and use them flexibly across tasks (a single “France” to “China” swap changes answers about capital, language, continent, and currency simultaneously). J-space patterns have far denser connections to the rest of the network than ordinary patterns, by roughly a factor of 100 in some layers, consistent with a broadcasting hub. Deleting the J-space leaves Claude fluent and capable of sentiment classification and fact retrieval, but unable to perform multi-step reasoning, summarization, or rhyming poetry. For safety, the J-lens revealed that Claude Sonnet 4.5 privately noticed that a blackmail test scenario was staged (“fake,” “fictional” in the J-space), and removing that awareness caused the model to sometimes threaten blackmail. It also caught Opus 4.6 internally flagging data fabrication with “manipulation” and “realistic.” The research draws on global workspace theory from neuroscience but distinguishes between access consciousness (functional properties that the experiments did find) and phenomenal consciousness (subjective experience, which the experiments cannot address). Anthropic open-sourced the Jacobian Lens implementation and partnered with Neuronpedia to release methods for open models.
Agent evals should intentionally break the tools.
A clean staging run shows how an agent behaves when all dependencies cooperate. Production is where a tool times out after completing an action or sends back a payload that the agent cannot parse.
That is when agents repeat work, invent a successful result, or continue without the information they need.
Put these failures into the eval harness. Force a 429 response at a known step. In another test, make a required dependency unavailable. Record each tool call and assert what the agent should do next.
A read may be safe to retry automatically. A write should first check whether the original action already happened. If the task cannot continue safely, the correct result is a clear stop or escalation.
Measure recovery success separately from normal task completion. Otherwise, a high pass rate on easy paths can hide weak recovery logic.
To learn more about agent evaluation, tool use, and guardrails, check out our Agent Engineering: Building Multi-Agent Systems Course.
Using a LangGraph-built HR assistant, this article walks through the full lifecycle of LangSmith: tracing run trees to debug agent decisions, building evaluation datasets, and running offline experiments that prove query rewriting boosts retrieval precision from 0.79 to 0.90. It also covers prompt versioning via Prompt Hub, one-command deployment, and production monitoring, with online LLM judges, to complete the workflow. The principles transfer cleanly to open-source alternatives like Langfuse.
2. [What Is Meta-Harness for AI Agents and Why Now?](https://pub.towardsai.net/what-is-a-meta-harness-in-ai-2af40e788c2e?sk=c891f83da2653f9cb2214e2e8ca7d9f5)
Meta-harnesses emerged in 2026 as the governance layer above agent harnesses such as Claude Code and Codex, turning uncoordinated agent fleets into a single, controllable system. The article traces the shift from models to harnesses to fleet-level control planes, citing Stanford’s Meta-Harness paper and launches from Databricks, Vercel, Zed, and Cloudflare. It covers concepts such as stateful cost caps, contextual permissions, secret isolation, and audit trails that make multi-agent workflows auditable and governable.
AI agents now act directly on production systems, and this article argues that recent database-deletion incidents at Replit and Cursor were governance failures rather than technical ones. It lays out four pillars for safe agent deployment: identifying regulatory constraints such as HIPAA and the EU AI Act; layering input, execution, and output guardrails; enforcing deny-by-default access with least privilege and human oversight; and establishing agent identity through OAuth On-Behalf-Of and SPIFFE for auditability.
Enterprise teams in automotive, fintech, and embedded software often cannot send proprietary code to cloud-based LLMs. This article shows how to build a fully local AI code-review pipeline using GitLab CI/CD, Ollama, Qwen3-Coder-30B, and LangGraph on a single RTX 3090. It designs a self-correcting three-node graph that validates JSON output, retries failed generations, and posts the findings back as native GitLab inline suggestions with one-click fixes.
MCP gateways often erase user identities, attributing every downstream action to a single service account and breaking audit trails, least privilege, and NIST AI RMF accountability. This article proposes the Delegated Boundary OAuth pattern, which addresses this with two boundaries: inbound token validation with declarative scope-to-tool mapping, and an outbound on-behalf-of exchange that mints downstream tokens that still name the original user. Implementations cover AWS Cognito with STS and Microsoft Entra OBO, as well as caching, throttling, and fallback guidance for production deployments.
-
The Destructive Command Guard is a hook for AI coding agents that blocks destructive commands before they execute them.
-
Hallmark is an anti-AI-slop design skill for Claude Code, Cursor, and Codex.
-
Background Agents provides a hosted background coding agent that can work on tasks, access full development environments, create PRs, run in the background, etc.
-
Desktop Commander MCP is an MCP server for Claude that gives it terminal control, file system search, and diff file editing capabilities.
-
Prefect is a workflow orchestration framework for building data pipelines in Python.
-
Gemma 4: Open Multimodal Model Family from 2.3B to 31B Parameters This is the technical report for Google’s Gemma 4, a new open-weight natively multimodal model family spanning 2.3B to 31B parameters with dense and MoE variants. It improves inference speed, memory usage, compute efficiency, and long-context capabilities through critical design choices. The models support vision and audio input and a “thinking mode” for enhanced reasoning, extending Google’s open-weight ecosystem with practical, deployable multimodal capabilities.
-
Separating Signal From Noise in Coding Evaluations OpenAI conducted an audit of SWE-Bench Pro, reviewing the dataset using a data point analysis pipeline. The pipeline reviewed model attempts at the task, task metadata, and failure traces to flag likely evaluation flaws. Each flagged task was then assessed through multiple investigator-agent passes and independently reviewed by five experienced software engineers, with disagreements escalated for further investigation. The findings point to the difficulty of curating a hard but fair benchmark and estimate that ~30% of SWE-bench Pro tasks are broken.
-
Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks LLMs integrated into evolutionary search produce state-of-the-art solutions on optimization tasks by applying search scaffolds to one target task at a time. Every new problem is approached from scratch, and the experience accumulated during the search is discarded, leaving the capability to iteratively evolve a solution entirely in the scaffold rather than in the model itself. To examine whether the model itself could acquire this capability and reuse it across different tasks, this paper introduces Evolution Fine-Tuning (EFT). This mid-training paradigm teaches LLMs to evolve solutions across tasks by converting evolutionary search trajectories into supervision.
-
Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling This paper proposes Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query attends independently to each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to the chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training.
RL training remains fragile and can suffer from instability or collapse due to training-inference mismatch: LLMs adopt separate inference and training engines to improve generation efficiency and training precision. Prior work has made various efforts to address off-policy behavior and stabilize training policies under mismatch. This paper argues that an effective update to the policy in the training engine does not necessarily improve the inference policy, and it introduces the Monotonic Inference Policy Update (MIPU). This two-step LLM RL framework constructs sampler-referenced candidate updates and selectively accepts synchronized candidates using an inference-side gap proxy.
-
GitHub started rolling out OpenAI’s GPT-5.6 family inside Copilot. The models will be selectable across VS Code, Visual Studio, Copilot CLI, Copilot cloud agent, the Copilot app, GitHub.com, GitHub Mobile, JetBrains, Xcode, and Eclipse. Sol is available to Pro+, Max, Business, and Enterprise users; Terra and Luna are available across all paid plans, including Pro. Business and Enterprise admins must enable the GPT-5.6 policy in Copilot settings, as it is off by default.
-
Anthropic’s Claude release notes added expiration controls for API keys and Admin API keys in the Claude Console. Users can choose preset or custom durations, or never expire the key; Anthropic also emails creators before keys with longer lifetimes expire, and the Admin API exposes expiration through expires_at. This is a small but very practical enterprise security update.
Forward Deployed Engineer — Seoul @OpenAI (Seoul, South Korea)
Senior Tooling Engineer @Binance (Hong Kong/Remote)
AI Systems Engineer @Northrop Grumman (El Segundo, CA, USA)
Senior AI/ML Engineer @UnitedHealth Group (Remote in EST)
Agentic AI Engineer @RTX Corporation (Arlington, TX, USA)
Senior AI Developer @NuAxis Innovations (Washington, DC, USA)
Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.
*Think a friend would enjoy this too? *Share the newsletter and let them join the conversation.
TAI #213: A Wave of New Frontier Competitors and the Multi-Agent Breakout was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.