Claude Managed Agents ramp up, Firebird opens its largest AI factory in the CIS, Kimi K3 returns to Copilot

On August 8, 2026, Anthropic expands Claude Managed Agents with a session budget, compute-region selection, automatic skill loading, and a consultable advisor model during the session. On the same day, Firebird inaugurates in Armenia the largest AI factory in the CIS region, with more than 70,000 GPUs targeted by the end of 2027, while GitHub confirms the return of Kimi K3 in Copilot with its exact pricing, generalizes the Lite and Balanced levels for code review, and publishes its weekly recap. Completing the picture: Warp Agent CLI, which connects an X Premium subscription to Grok; Replit’s internal assessment of massive agent usage; Meta Muse Spark 1.2, which confirms its competitiveness on two rankings; and about fifteen briefs from Hugging Face, Together AI, and generative media.

Claude Developer Platform: four updates for Managed Agents

August 8 — Anthropic delivered four improvements to Claude Managed Agents on the Claude Developer Platform. A budget can now be set per session to control spending: a session that reaches the limit pauses with a budget_reached event, and raising the budget is enough to resume. The model.inference_geo parameter accepts us (billed at 1.1x, US region guaranteed) or global (standard pricing, inference wherever capacity is available). Managed Agents also automatically load skills already present in .claude/skills/ from attached repositories, with no extra configuration. Finally, a Managed Agent can now receive a more powerful “advisor” model, available during the session for a second opinion — a single roster configuration line is enough.

Update	Detail
Session budget	Automatic pause at the limit, `budget_reached` event
Compute region	`us` (1.1x) or `global` (standard pricing)
Skill loading	Automatic from attached repositories
Advisor model	Second opinion during the session, 1 configuration line

🔗 Announcement @ClaudeDevs

Coding tools: Warp connects X Premium, Replit publishes its internal agent report

Warp Agent CLI adds /connect-grok to link an X Premium subscription

August 7 — Warp adds the /connect-grok command to the Warp Agent CLI, launched on August 4, to link an existing X Premium subscription and use Grok without configuring a separate API key — an extension of the “bring your own key” model already offered by the tool. The same day, Warp published a video demonstration showcasing several features together: multi-agent orchestration, plan mode with syntax highlighting of reasoning, a voice mode (/voice), and a shell mode.

Run /connect-grok in the Warp Agent CLI to use your X Premium subscription with Warp! — @warpdotdev on X

Replit details how its teams almost tripled their code output with agents

August 7 — Replit publishes an internal report on 6 months of using its own agents internally: the amount of code produced has almost tripled, while review time, regression rate, and incidents have remained stable, and quality has improved. The team stresses that this is not a single AI but an agent system distributed across several functions — incident investigation, pull request review, support triage, data analysis, sales account research. Another notable point: a paid SaaS tool used internally was replaced by an application built entirely on Replit, judged better than the original. Replit describes this evolution as the beginning of a “self-piloted company.”

Over the last 6 months, Replit nearly tripled its code output. Quality held. Nothing broke. We started becoming a self-driving company. — @Replit on X

Meta Muse Spark 1.2 confirms its competitiveness on two rankings

August 6 — Two independent rankings confirm the breakthrough of Meta Muse Spark 1.2. On the Vals Index (general ranking), the model enters the top 5 at $0.69 per test — about 3 times cheaper than Kimi and 10 times cheaper or more than Opus, Fable, or GPT-5.6 Sol for a comparable position. On a ranking specialized in video game development, Muse Spark 1.2 matches GPT-5.6-Sol and ranks tied for 3rd, a position a Meta team engineer says would have been unattainable at the start of the year for the company on this type of ranking. These two signals add to the gold medals at the International Science Olympiads and the Artificial Analysis benchmark already covered earlier in August, confirming a continuous upward trend in the competitiveness of the Muse Spark family, with a clear cost advantage.

Ranking	Muse Spark 1.2 position	Reference
Vals Index	Top 5	$0.69/test, 3x cheaper than Kimi
Video games (specialized)	Tied for 3rd	Matches GPT-5.6-Sol

Muse Spark 1.2 just cracked the top 5 on the Vals Index, at just $0.69 per test. This is 3x cheaper than Kimi and 10x or more cheaper than Fable, Opus, and 5.6 Sol. — @ValsAI on X

GitHub Copilot: Kimi K3, weekly updates, and code review

Kimi K3 in Copilot: deployment paused then resumed, pricing confirmed

August 6-7 — GitHub pauses the rollout of Kimi K3 in Copilot on August 6 while resolving an incident affecting GitHub Actions, while also communicating the model’s planned pricing in advance. The next day, August 7, GitHub announces the rollout has resumed: Kimi K3 is billed at the provider list pricing under usage-based billing, and the company points to its pricing documentation for the full details of models and requests. This episode completes the general availability of Kimi K3 in Copilot announced the day before: it adds the exact pricing information and the technical incident details that briefly interrupted the rollout.

Token type	Price
Input	$3 / million
Output	$15 / million
Cached input	$0.30 / million

🔗 Tweet about rollout resuming

Weekly updates for August 3: app, CLI, VS Code

August 7 — GitHub Copilot’s weekly recap covers the desktop app, CLI, and VS Code 1.132. In the desktop app, Copilot now displays the model used per request with details of consumed credits and cache usage, and allows users to join shared sessions or open a parallel exploration via /side. The CLI handles multiple simultaneous sessions in a sidebar (shortcuts n to create, x to navigate), adds an experimental /worktree command for an isolated workspace and a /rewind command without Git dependency to restore a previous state. In VS Code 1.132, multilingual dictation automatically detects the spoken language, and the /btw command lets users ask a follow-up question without interrupting the ongoing conversation.

Surface	Key update
Desktop app	Model and credits shown per request, `/side` command
CLI	Multiple sessions, `/worktree`, `/rewind` without Git
VS Code 1.132	Multilingual dictation, `/btw` command

🔗 August 3 weekly updates

Copilot code review: Lite and Balanced move to general availability

August 7 — GitHub generalizes two effort levels for Copilot code review, replacing the old Low and Medium designations. The Lite level targets simple changes (documentation, minor fixes) with focused feedback; the Balanced level is for complex logic, security-sensitive code, and cross-service changes, with deeper analysis relying on strengthened reasoning models. The choice is made per review without changing repository or organization defaults, but administrators can define a default level applicable to all repositories from organization settings. The selected level now appears in the pull request timeline and summary comment. Both levels are included in Copilot Pro, Pro+, Max, Business, and Enterprise.

Level	Target	Included plans
Lite	Simple changes, docs, minor fixes	Pro, Pro+, Max, Business, Enterprise
Balanced	Complex logic, security, cross-service	Pro, Pro+, Max, Business, Enterprise

🔗 Copilot code review effort levels are generally available

Firebird launches the largest AI factory in the CIS region in Armenia

August 8 — Firebird, an emerging AI cloud active in the CIS region (Commonwealth of Independent States), inaugurated in Hrazdan, Armenia, what is presented as the region’s largest AI factory. The infrastructure relies on NVIDIA accelerated computing platform and Dell Technologies PowerEdge servers, with a NVIDIA DSX platform claiming 40% more GPUs in the same footprint thanks to its energy efficiency. The deployment plans for more than 70,000 GPUs (a mix of Blackwell and Rubin) by the end of 2027, for a target capacity of 300 megawatts, brought online in just over six months. Perplexity is cited as an early customer. Firebird announces a regional 2-gigawatt roadmap, with Armenia and Kazakhstan as the first milestones.

Item	Detail
Deployed GPUs	70,000+ (Blackwell + Rubin) by end of 2027
Target capacity	300 MW
Partners	NVIDIA, CoreWeave, Dell Technologies, Schneider Electric, Vertiv
Early customer	Perplexity

🔗 Firebird Launches CIS Region’s Largest AI Factory in Armenia

Briefs

Replit lets users move a project between workspaces — Transfer possible between workspaces on the same team or enterprise account, from the project list dropdown menu. 🔗 source
New DNA foundation model published on Hugging Face — A model capable of reading and generating DNA sequences is available with a demo space; model name and organization not specified. 🔗 source
NVIDIA publishes NeMo Gym conversational tool-use assets — Reference policy/tool pairs and prompt histories published on Hugging Face to train conversational agents that use tools. 🔗 source
Hugging Face shares a study on AI agents reproducing ICML 2026 papers — Thread on the automated reproduction of scientific publications by AI agents, with no methodological details or success rates. 🔗 source
Hugging Face Storage Buckets integrated into Vast AI — GPU instances rented on Vast AI can now read and write directly to a Hugging Face bucket, without manual transfer. 🔗 source
A third-party provider claims the fastest inference for DeepSeek V4 Flash and Kimi K3 — Claim based on real production traffic rather than synthetic tests, to be taken with caution. 🔗 source
Together AI adds a Learn section to its documentation — Educational content explaining time to first token (TTFT), context windows, sampling, fine-tuning, quantization, and deployment trade-offs. 🔗 source
FLUX 3 available in production on Together AI — The multimodal model from Black Forest Labs (video and synchronized audio, clips up to 20 seconds) is accessible via Together AI’s managed serverless inference. 🔗 source
Roomote chooses Together AI as its inference provider — The agent tool can now assign different open models depending on the workflow stage: code, planning, vision, and review. 🔗 source
The Copilot impact dashboard adds a return on investment section — New section comparing the monthly cost of Copilot per developer to the number of pull requests produced, with an interactive salary selector. 🔗 source
Companies can now install third-party GitHub Apps — Enterprise account owners can install external public GitHub Apps, with limited access to the enterprise account and restricted sensitive permissions. 🔗 source
MiniMax H3 number 1 in three DesignArena categories — The open weights model claims first place in Multi-Image to Video, Image to Video, and Video Editing on the DesignArena Frontier ranking. 🔗 source
Kimi K3 — strong growth in downloads and active users (SensorTower) — The Kimi app’s downloads have nearly quintupled and daily active users have jumped by about 40% since the launch of Kimi K3. 🔗 source

What this means

Agentic infrastructure is becoming standardized among the major providers. Anthropic is adding to Managed Agents what Claude Code already offered on its own — budget, region selection, skill loading — and is innovating with a consultable adviser model during the session. GitHub is extending the same logic on the governance side: Copilot code review is rolling out Lite and Balanced as organization-level configurations, while the CLI is adding reversible commands (/rewind) that treat the session as a state to restore rather than a disposable stream. Warp, for its part, is broadening access to third-party models (Grok via an X Premium subscription) without forcing a dedicated API key. The common thread: agentic tools are adopting the same reflexes as traditional cloud infrastructure — budgets, regions, permissions, rollback.

Cost/performance continues to dictate adoption choices more than raw rankings. Meta Muse Spark 1.2 enters the top 5 of the Vals Index at $0.69 per test, 3 times cheaper than Kimi despite a comparable positioning; Kimi K3, for its part, sees its downloads almost quintuple according to SensorTower while also confirming precise usage-based pricing on GitHub Copilot ($3 / $15 per million tokens). Together, these two signals sketch a market where user traction and cost per request weigh just as much as placement on a reference leaderboard.

Physical infrastructure deployment continues to accelerate outside traditional zones. Firebird chooses Armenia to build the largest AI factory in the CIS region, with more than 70,000 GPUs targeted by the end of 2027 and a regional 2-gigawatt roadmap. Perplexity, cited as an early customer, illustrates a movement in which model providers are actively seeking geographic diversity in compute rather than relying solely on U.S. or Chinese hubs.

Finally, the satellite tools ecosystem is consolidating around a handful of open inference providers. Together AI is chaining together integrations in one week — FLUX 3 in production, Roomote as an agentic client, a dedicated educational section on inference concepts — while Hugging Face is expanding its role as a storage platform (Vast AI integration) as much as a model distribution platform. This quiet movement, made up of brief updates rather than spectacular announcements, is building the plumbing on which today’s more visible uses rest.