|
Sr. Content Developer at Microsoft, working remotely in PA, TechBash conference organizer, former Microsoft MVP, Husband, Dad and Geek.
|
We’re excited to announce the release of Pulumi Kubernetes Operator v2.3.0, introducing two powerful capabilities that enhance GitOps workflows: preview mode for validating infrastructure changes before deployment, and structured configuration support for managing complex data types. Building on the success of the v2.0 GA release, this update addresses long-standing community requests while maintaining full backwards compatibility. These features enable safer, more sophisticated infrastructure management patterns for platform engineering teams.
Preview mode enables you to run Pulumi stacks in dry-run fashion, allowing you to visualize what infrastructure changes would occur without actually applying them. This capability is essential for GitOps workflows that require validation gates and incremental rollouts.
By adding preview: true to your Stack specification, the operator runs pulumi preview instead of pulumi up. The Stack’s Ready condition indicates preview success, and you get full status including preview links, standard output, and program outputs—all without making actual infrastructure changes.
This unlocks sophisticated workflow patterns:
Log in
preview flag on and off with external verification steps between each deployment phaseHere’s a simple example showing a production stack alongside its preview counterpart:
# Production stack
apiVersion: pulumi.com/v1
kind: Stack
metadata:
name: my-infrastructure
spec:
stack: org/project/prod
projectRepo: https://github.com/example/infra
branch: main
---
# Preview stack - same Pulumi stack, preview mode
apiVersion: pulumi.com/v1
kind: Stack
metadata:
name: my-infrastructure-preview
spec:
stack: org/project/prod
projectRepo: https://github.com/example/infra
branch: feature-branch
preview: true # Only runs pulumi preview
Preview mode closes issue #16, one of our longest-standing feature requests.
Configuration management takes a significant step forward with native support for complex data types. Previously limited to string values, Stack configuration now supports objects, arrays, numbers, and booleans as first-class citizens.
This enhancement addresses the reality that complex environments require rich configuration. You can now express sophisticated configuration structures inline in your Stack manifests or load them from ConfigMaps with automatic JSON parsing—all using Kubernetes-native patterns.
The implementation leverages Pulumi CLI’s JSON configuration support (available in v3.202.0+) with automatic version detection. If your workspace uses an earlier CLI version, you’ll receive clear guidance to upgrade. Existing string-only configurations continue to work without modification, ensuring full backwards compatibility.
Here’s an example of inline structured configuration:
apiVersion: pulumi.com/v1
kind: Stack
metadata:
name: my-app
spec:
stack: org/app/prod
projectRepo: https://github.com/example/app
branch: main
config:
# String values (existing behavior)
environment: "production"
# Objects (NEW)
database:
host: "db.example.com"
port: 5432
ssl: true
# Arrays (NEW)
regions: ["us-west-2", "us-east-1", "eu-west-1"]
# Numbers and booleans (NEW)
maxConnections: 100
enableCaching: true
You can also reference ConfigMaps for more complex scenarios:
apiVersion: v1
kind: ConfigMap
metadata:
name: app-settings
data:
database.json: |
{
"host": "db.example.com",
"port": 5432,
"maxConnections": 100
}
---
apiVersion: pulumi.com/v1
kind: Stack
metadata:
name: my-app
spec:
stack: org/app/prod
projectRepo: https://github.com/example/app
branch: main
configRefs:
database:
name: app-settings
key: database.json
json: true # Parse as JSON
Note that Secrets are not a supported source of structured configuration values.
Structured configuration closes issue #258 and issue #872, addressing long-standing configuration management needs from the community.
This release includes several fixes that improve operational reliability:
secretsProvider parameter now properly applies when creating new stacks (#935)podLabels containing special characters like colons (#1014)ttlAfterCompleted setting for automatic garbage collection (#960)This release includes updates to the Custom Resource Definitions (CRDs) to support the new features. If you’re using Helm, you’ll need to manually apply the updated CRDs before upgrading, as Helm v3 does not automatically upgrade CRDs:
# Apply updated CRDs
kubectl apply --server-side -k 'github.com/pulumi/pulumi-kubernetes-operator//operator/config/crd?ref=v2.3.0'
# Upgrade via Helm
helm upgrade pulumi-kubernetes-operator \
oci://ghcr.io/pulumi/helm-charts/pulumi-kubernetes-operator \
--version 2.3.0 \
-n pulumi-kubernetes-operator
If you’re using the quickstart YAML installation method, the CRDs will update automatically when you apply the new manifest.
All changes are backwards compatible—existing Stack resources work without modification. For structured configuration support, ensure your workspace pods use Pulumi CLI v3.202.0 or later; the operator provides automatic version detection with clear upgrade guidance when needed.
The v2.3.0 release enhances the Pulumi Kubernetes Operator with safer deployment workflows and more flexible configuration management. Preview mode enables validation-first GitOps patterns, while structured configuration simplifies complex multi-environment setups.
Get started with the Pulumi Kubernetes Operator in our documentation, or view the complete v2.3.0 release notes on GitHub. Join the conversation in the Pulumi Community Slack #kubernetes channel—we’d love to hear how these features impact your workflows.
Chinese artificial intelligence startup DeepSeek released two powerful new AI models on Sunday that the company claims match or exceed the capabilities of OpenAI's GPT-5 and Google's Gemini-3.0-Pro — a development that could reshape the competitive landscape between American tech giants and their Chinese challengers.
The Hangzhou-based company launched DeepSeek-V3.2, designed as an everyday reasoning assistant, alongside DeepSeek-V3.2-Speciale, a high-powered variant that achieved gold-medal performance in four elite international competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, the ICPC World Finals, and the China Mathematical Olympiad.
The release carries profound implications for American technology leadership. DeepSeek has once again demonstrated that it can produce frontier AI systems despite U.S. export controls that restrict China's access to advanced Nvidia chips — and it has done so while making its models freely available under an open-source MIT license.
"People thought DeepSeek gave a one-time breakthrough but we came back much bigger," wrote Chen Fang, who identified himself as a contributor to the project, on X (formerly Twitter). The release drew swift reactions online, with one user declaring: "Rest in peace, ChatGPT."
At the heart of the new release lies DeepSeek Sparse Attention, or DSA — a novel architectural innovation that dramatically reduces the computational burden of running AI models on long documents and complex tasks.
Traditional AI attention mechanisms, the core technology allowing language models to understand context, scale poorly as input length increases. Processing a document twice as long typically requires four times the computation. DeepSeek's approach breaks this constraint using what the company calls a "lightning indexer" that identifies only the most relevant portions of context for each query, ignoring the rest.
According to DeepSeek's technical report, DSA reduces inference costs by roughly half compared to previous models when processing long sequences. The architecture "substantially reduces computational complexity while preserving model performance," the report states.
Processing 128,000 tokens — roughly equivalent to a 300-page book — now costs approximately $0.70 per million tokens for decoding, compared to $2.40 for the previous V3.1-Terminus model. That represents a 70% reduction in inference costs.
The 685-billion-parameter models support context windows of 128,000 tokens, making them suitable for analyzing lengthy documents, codebases, and research papers. DeepSeek's technical report notes that independent evaluations on long-context benchmarks show V3.2 performing on par with or better than its predecessor "despite incorporating a sparse attention mechanism."
DeepSeek's claims of parity with America's leading AI systems rest on extensive testing across mathematics, coding, and reasoning tasks — and the numbers are striking.
On AIME 2025, a prestigious American mathematics competition, DeepSeek-V3.2-Speciale achieved a 96.0% pass rate, compared to 94.6% for GPT-5-High and 95.0% for Gemini-3.0-Pro. On the Harvard-MIT Mathematics Tournament, the Speciale variant scored 99.2%, surpassing Gemini's 97.5%.
The standard V3.2 model, optimized for everyday use, scored 93.1% on AIME and 92.5% on HMMT — marginally below frontier models but achieved with substantially fewer computational resources.
Most striking are the competition results. DeepSeek-V3.2-Speciale scored 35 out of 42 points on the 2025 International Mathematical Olympiad, earning gold-medal status. At the International Olympiad in Informatics, it scored 492 out of 600 points — also gold, ranking 10th overall. The model solved 10 of 12 problems at the ICPC World Finals, placing second.
These results came without internet access or tools during testing. DeepSeek's report states that "testing strictly adheres to the contest's time and attempt limits."
On coding benchmarks, DeepSeek-V3.2 resolved 73.1% of real-world software bugs on SWE-Verified, competitive with GPT-5-High at 74.9%. On Terminal Bench 2.0, measuring complex coding workflows, DeepSeek scored 46.4%—well above GPT-5-High's 35.2%.
The company acknowledges limitations. "Token efficiency remains a challenge," the technical report states, noting that DeepSeek "typically requires longer generation trajectories" to match Gemini-3.0-Pro's output quality.
Beyond raw reasoning, DeepSeek-V3.2 introduces "thinking in tool-use" — the ability to reason through problems while simultaneously executing code, searching the web, and manipulating files.
Previous AI models faced a frustrating limitation: each time they called an external tool, they lost their train of thought and had to restart reasoning from scratch. DeepSeek's architecture preserves the reasoning trace across multiple tool calls, enabling fluid multi-step problem solving.
To train this capability, the company built a massive synthetic data pipeline generating over 1,800 distinct task environments and 85,000 complex instructions. These included challenges like multi-day trip planning with budget constraints, software bug fixes across eight programming languages, and web-based research requiring dozens of searches.
The technical report describes one example: planning a three-day trip from Hangzhou with constraints on hotel prices, restaurant ratings, and attraction costs that vary based on accommodation choices. Such tasks are "hard to solve but easy to verify," making them ideal for training AI agents.
DeepSeek employed real-world tools during training — actual web search APIs, coding environments, and Jupyter notebooks — while generating synthetic prompts to ensure diversity. The result is a model that generalizes to unseen tools and environments, a critical capability for real-world deployment.
Unlike OpenAI and Anthropic, which guard their most powerful models as proprietary assets, DeepSeek has released both V3.2 and V3.2-Speciale under the MIT license — one of the most permissive open-source frameworks available.
Any developer, researcher, or company can download, modify, and deploy the 685-billion-parameter models without restriction. Full model weights, training code, and documentation are available on Hugging Face, the leading platform for AI model sharing.
The strategic implications are significant. By making frontier-capable models freely available, DeepSeek undermines competitors charging premium API prices. The Hugging Face model card notes that DeepSeek has provided Python scripts and test cases "demonstrating how to encode messages in OpenAI-compatible format" — making migration from competing services straightforward.
For enterprise customers, the value proposition is compelling: frontier performance at dramatically lower cost, with deployment flexibility. But data residency concerns and regulatory uncertainty may limit adoption in sensitive applications — particularly given DeepSeek's Chinese origins.
DeepSeek's global expansion faces mounting resistance. In June, Berlin's data protection commissioner Meike Kamp declared that DeepSeek's transfer of German user data to China is "unlawful" under EU rules, asking Apple and Google to consider blocking the app.
The German authority expressed concern that "Chinese authorities have extensive access rights to personal data within the sphere of influence of Chinese companies." Italy ordered DeepSeek to block its app in February. U.S. lawmakers have moved to ban the service from government devices, citing national security concerns.
Questions also persist about U.S. export controls designed to limit China's AI capabilities. In August, DeepSeek hinted that China would soon have "next generation" domestically built chips to support its models. The company indicated its systems work with Chinese-made chips from Huawei and Cambricon without additional setup.
DeepSeek's original V3 model was reportedly trained on roughly 2,000 older Nvidia H800 chips — hardware since restricted for China export. The company has not disclosed what powered V3.2 training, but its continued advancement suggests export controls alone cannot halt Chinese AI progress.
The release arrives at a pivotal moment. After years of massive investment, some analysts question whether an AI bubble is forming. DeepSeek's ability to match American frontier models at a fraction of the cost challenges assumptions that AI leadership requires enormous capital expenditure.
The company's technical report reveals that post-training investment now exceeds 10% of pre-training costs — a substantial allocation credited for reasoning improvements. But DeepSeek acknowledges gaps: "The breadth of world knowledge in DeepSeek-V3.2 still lags behind leading proprietary models," the report states. The company plans to address this by scaling pre-training compute.
DeepSeek-V3.2-Speciale remains available through a temporary API until December 15, when its capabilities will merge into the standard release. The Speciale variant is designed exclusively for deep reasoning and does not support tool calling — a limitation the standard model addresses.
For now, the AI race between the United States and China has entered a new phase. DeepSeek's release demonstrates that open-source models can achieve frontier performance, that efficiency innovations can slash costs dramatically, and that the most powerful AI systems may soon be freely available to anyone with an internet connection.
As one commenter on X observed: "Deepseek just casually breaking those historic benchmarks set by Gemini is bonkers."
The question is no longer whether Chinese AI can compete with Silicon Valley. It's whether American companies can maintain their lead when their Chinese rival gives comparable technology away for free.
As we make our way to AWS re:Invent today in Las Vegas, these are some of the questions on our mind: Will Amazon CEO Andy Jassy make another appearance? Will this, in fact, be Amazon CTO Werner Vogels’ last big closing keynote at the event? Will we be able to line up early enough to score a seat inside the special Acquired podcast recording Thursday morning?
And how many million enterprise AI billboards will we see between the airport and the Venetian?
But more to the point for Amazon, the company faces a critical test this week: showing that its heavy artificial intelligence investments can pay off as Microsoft and Google gain ground in AI and the cloud.
A year after the Seattle company unveiled its in-house Nova AI foundation models, the expansion into agentic AI will be the central theme as Amazon Web Services CEO Matt Garman takes the stage Tuesday morning for the opening keynote at the company’s annual cloud conference.
The stakes are big, for both the short and long term. AWS accounts for a fifth of Amazon’s sales and more than half of its profits in many quarters, and all the major cloud platforms are competing head-to-head in AI as the next big driver of growth.
With much of the tech world focused on the AI chip race, the conference will be closely watched across the industry for news of the latest advances in Amazon’s in-house Trainium AI chips.
But even as the markets and outside observers focus on AI, we’ve learned from covering this event over the years that many AWS customers still care just as much or more about advances in the fundamental building blocks of storage, compute and database services.
Amazon gave a hint of its focus in early announcements from the conference:
Beyond the product news, AWS is making a concerted effort to show that the AI boom isn’t just for the big platforms. In a pitch to consultants and integrators at the event, the company released new research from Omdia, commissioned by Amazon, claiming that partners can generate more than $7 in services revenue for every dollar of AWS technology sold.
Along with that research, AWS launched a new “Agentic AI” competency program for partners, designed to recognize firms building autonomous systems rather than simple chatbots.
Garman’s keynote begins at 8 a.m. PT Tuesday, with a dedicated agentic AI keynote from VP Swami Sivasubramanian on Wednesday, an infrastructure keynote on Thursday morning, and Vogels’ aforementioned potential swan song on Thursday afternoon.
Stay tuned to GeekWire for coverage, assuming we make it to the Strip!
We recently shipped Agent HQ’s mission control, a unified interface for managing GitHub Copilot coding agent tasks.
Now, you can now assign tasks to Copilot across repos, pick a custom agent, watch real‑time session logs, steer mid-run (pause, refine, or restart), and jump straight into the resulting pull requests—all in one place. Instead of bouncing between pages to see status, rationale, and changes, mission control centralizes assignment, oversight, and review.
Having the tool is one thing. Knowing how to use it effectively is another. This guide shows you how to orchestrate multiple agents, when to intervene, and how to review their work efficiently. Being great at orchestrating agents means unblocking parallel work in the same timeframe you’d spend on one task, stepping in when logs show drift, tests fail, or scope creeps.
If you’re already used to working with an agent one at a time, you know it’s inherently sequential. You submit a prompt, wait for a response, review it, make adjustments, and move to the next task.
Mission control changes this. You can kick off multiple tasks in minutes—across one repo or many. Previously, you’d navigate to different repos, open issues in each one, and assign Copilot separately. Now you can enter prompts in one place, and Copilot coding agent goes to work across all of them.
That being said, there is a trade-off to keep in mind: Instead of each task taking30 seconds to a few minutes to complete, your agents might spend a few minutes to an hour on a draft. But you’re no longer just waiting. You’re orchestrating.
Not everything belongs in parallel. Use sequential workflows when:
When assigning multiple tasks from the same repo, consider overlap. Agents working in parallel can create merge conflicts if they touch the same files. Be thoughtful about partitioning work.
Tasks that typically run well in parallel:
The shift is simple: you move from waiting on a single run to overseeing multiple progressing in parallel, stepping in for failed tests, scope drift, or correcting unclear intent where guidance will save time.
Specificity matters. Describe the task precisely. Good context remains critical for good results.
Helpful context includes:
Weak prompt: “Fix the authentication bug.”
Strong prompt: “Users report ‘Invalid token’ errors after 30 minutes of activity. JWT tokens are configured with 1-hour expiration in auth.config.js. Investigate why tokens expire early and fix the validation logic. Create the pull request in the api-gateway repo.”
Mission control lets you select custom agents that use agents.md files from your selected repo. These files give your agent a persona and pre-written context, removing the burden of constantly providing the same examples or instructions.
If you manage repos where your team regularly uses agents, consider creating agents.md files tailored to your common workflows. This ensures consistency across tasks and reduces the cognitive load of crafting detailed prompts each time.
Once you’ve written your prompt and selected your custom agent (if applicable), kick off the task. Your agent gets to work immediately.
You’re now a conductor of agents. Each task might take a minute or an hour, depending on complexity. You have two choices: watch your agents work so you can intervene if needed, or step away and come back when they’re done.
Below are some common indicators that your agent is not on the right track and needs additional guidance:
When you spot issues, evaluate their severity. Is that failing test critical? Does that integration point matter for this task? The session log typically shows intent before action, giving you a chance to intervene if you’re monitoring.
When you need to redirect an agent, be specific. Explain why you’re redirecting and how you want it to proceed.
Bad steering: “This doesn’t look right.”
Good steering: “Don’t modify database.js—that file is shared across services. Instead, add the connection pool configuration in api/config/db-pool.js. This keeps the change isolated to the API layer.”
Timing matters. Catch a problem five minutes in, and you might save an hour of ineffective work. Don’t wait until the agent finishes to provide feedback.
You can also stop an agent mid-task and give it refined instructions. Restarting with better direction is simple and often faster than letting a misaligned agent continue.
Session logs show reasoning, not just actions. They reveal misunderstandings before they become pull requests, and they improve your future prompts and orchestration practices. When Copilot says “I’m going to refactor the entire authentication system,” that’s your cue to steer.
When your agents finish, you’ll have pull requests to review. Here’s how to do it efficiently. Ensure you review:
This pattern gives you the full picture: intent, implementation, and validation.
After an agent completes a task, ask it:
Copilot can often identify gaps in its own work, saving you time and improving the final result. Treat it like a junior developer who’s willing to explain their reasoning.
Generating code with agents is straightforward. Reviewing that code—ensuring it meets your standards, does what you want, and that it can be maintained by your team—still requires human judgment.
Improve your review process by grouping similar work together. Review all API changes in one session. Review all documentation changes in another. Your brain context-switches less, and you’ll spot patterns and inconsistencies more easily.
Mission control moves you from babysitting single agent runs to orchestrating a small fleet. You define clear, scoped tasks. You supply just enough context. You launch several agents. The speed gain is not that each task finishes faster; it’s that you unblock more work in the same timeframe.
What makes this possible is discipline: specific prompts, not vague requests. Custom agents in agents.md that carry your patterns so you don’t repeat yourself. Early steering when session logs show drift. Treating logs as reasoning artifacts you mine to write a sharper next prompt. And batching reviews so your brain stays in one mental model long enough to spot subtle inconsistencies. Lead your own team of agents to create something great!
Ready to start? Visit mission control or learn more about GitHub Copilot for your organization.
The post How to orchestrate agents using mission control appeared first on The GitHub Blog.
