blog
writing about what we build
platform updates, industry research, and technical deep dives.
Seventy-Two Hours: What Claude Fable 5 Means for Agentic Ops
Anthropic shipped Claude Fable 5 — its most capable model, built for long-horizon agentic work — on June 9. On June 12 a U.S. government directive suspended it for every customer. The frontier raises what autonomous ops can attempt; the suspension proves why you never hardwire one model id. Route across providers, own a floor you control, keep identity and audit in your own platform.
Run as the user, not the robot: Entra OBO identity across AWS, GCP, and Kubernetes
Shared service accounts and standing keys make every privileged action anonymous — exactly what FedRAMP High and NIST 800-53 high-impact forbid. The fix is an identity model where the agent runs as the Entra user: on-behalf-of token exchange federates one human identity down into AWS STS, GCP Workload Identity, and Kubernetes RBAC, so CloudTrail, Cloud Audit Logs, and the k8s audit log all name the person — not a robot. The AWS path is implemented in the open-source core today; here is the architecture, control by control, and how to wire it.
Everyone claims it. No one ships it. The open-source core is real — and free.
The AI-SRE category finally ships real software — Komodor, Resolve.ai, Cleric, NeuBird. But almost all of it is rented SaaS that runs as a shared bot. Our open-source core is self-hosted, multi-provider, free, and runs as the human, on your own identity — with the receipts of it running v1.0.0 on a real Kubernetes cluster.
Above the Workflow: Fleet, Mission, and the Orchestration Layer We're Building for Enterprise AIOps
Workflows that coordinate many agents on one issue are real and running today. The layer above them — Missions that spawn many workflows toward one objective, and Fleets that federate deployments — is in active prototype. Here is the vision, and an honest map of the seam.
The Squeeze: How AI Providers Are Cornering Enterprise Teams — and the Open Ecosystem That Breaks Free
Token costs are rising 10–60x as flat-rate AI subscriptions end. Proprietary toolchains collect your behavioral data. Anthropic has released zero open-source models. Meanwhile, Gemma 4 runs on Ollama at frontier performance with 12.8M downloads. Here is what is happening and why the platform and the open-source core exist.
We Wrote 101 Use Cases Before We Wrote the Platform. Here's What That Bought Us.
Most AI products are solutions chasing problems — 80% of enterprise workers reject them, 80%+ of POCs never ship. The platform was engineered the other way: 101 numbered business cases first, platform capabilities second. Here is why that inversion ships outcomes instead of demos, and what's live in v0.6.5 "Lock Step".
80% of Workers Are Rejecting AI. The Problem Isn't the Technology — It's the Deployment Architecture.
A new WalkMe survey finds 80% of enterprise workers avoiding AI tools while companies spend $54M+ on digital transformation. Fortune's reporting reveals a Ferrari with no driver, no fuel, and no roads. Here's the architecture that fixes all three.
inside claude code: a complete technical analysis of anthropic's leaked cli source
On March 31, Anthropic accidentally shipped a source map in their npm package that exposed 513,000 lines of Claude Code's TypeScript source. We read every line. Here's what we found — the architecture, the telemetry, the permissions system, and what it means for enterprise AI.
Introducing GhostPilot: Because Developers Shouldn't Have to Write Playwright Scripts
GhostPilot lets human and AI share the same Chromium browser in real time. No CSS selectors, no DOM dumps. Natural language browser automation with secret redaction and enterprise audit trails.
AI Agents in Your Channels: Slack and Teams Integration
The platform brings live AI agents to Slack and Teams — trigger workflows, get real-time reports, and run autonomous operations from your messaging channels.
Anthropic Sues the Pentagon: A Legal Analysis of the Supply Chain Risk Designation
Anthropic filed two federal lawsuits challenging its supply chain risk designation. Legal analysis of the First Amendment, due process, and FASCSA arguments — and what this means for every AI company.
Why Synth Makes MCP Servers, OpenClaw, and Static Tool Registries Obsolete
Synth is an ephemeral, sandboxed tool execution model that eliminates the persistent attack surfaces of MCP servers, OpenClaw, and static tool registries.
Why We Built Our Platform: The Architecture Behind Enterprise-Grade Agentic AI
Deep technical breakdown of the platform's CodeMode, ChatMode, Workflow Engine, and SmartModelRouter — the architecture behind enterprise-grade agentic AI.
The Anthropic-Pentagon Crisis: What Government AI Dependency Really Costs — and How to Avoid It
The Anthropic-Pentagon ban proves why government AI infrastructure cannot depend on a single vendor. Technical analysis of how model-agnostic platforms eliminate this risk entirely.
AI Provider Data Collection: What Actually Gets Transmitted When You Send a Prompt
Exactly what gets transmitted when you send a prompt — verified against official documentation from Anthropic, AWS Bedrock, OpenAI, Azure AI Foundry, Google Vertex AI, and Ollama.
From Alert to Resolution in Minutes: Autonomous Incident Response with Our Platform
The platform resolves the vast majority of PagerDuty alerts autonomously using multi-agent parallel processing, scoped RBAC credentials, and on-demand tool synthesis. From alert to resolution in minutes, not hours.
ClawHavoc: The Supply Chain Attack That Proves AI Agent Registries Are Broken
ClawHavoc compromised 824+ skills in OpenClaw's ClawHub registry with credential-stealing malware. This is why the "app store for AI tools" model is broken and what the alternative looks like.
What Is OpenClaw? The Autonomous AI Agent That Proves Your Security Model Is Broken
OpenClaw is the fastest-growing open-source AI agent — and a case study in why autonomous agents need governance. Here is what it gets right, what it gets wrong, and why enterprises need our platform.
AI Supply Chain Attacks Are Here: PyTorch, Ultralytics, and What Comes Next
AI supply chain attacks are no longer theoretical. PyTorch, Ultralytics, and Hugging Face have all been compromised. Here is why enterprise AI needs governed agent operations.
Why 40% of Agentic AI Projects Will Fail — And How Synth Changes the Math
Gartner predicted 40%+ of agentic AI projects will fail. Here are the four failure modes killing enterprise AI — and how the Synth framework structurally prevents each one.
MCP Is the USB of AI — But You Still Need an Electrician
MCP hit 97M monthly SDK downloads and joined the Linux Foundation. It is the universal connector for AI agents — but production enterprise deployments need security, governance, and operations layers that MCP alone does not provide.
Tool Synthesis, Not Tool Registration: How Synth Eliminates AI Integration Debt
Traditional agentic frameworks force you to pre-build tool wrappers for every API. Synth synthesizes tools on demand, eliminating integration debt entirely.
94% of IT Leaders Fear Vendor Lock-In. Here's the Architecture That Prevents It.
A Parallels survey found 94% of IT leaders fear vendor lock-in. AI lock-in is deeper than SaaS — spanning models, infrastructure, frameworks, and data. Here are the architectural patterns that prevent it.
Human-in-the-Loop AI: From Buzzword to Architecture Pattern
Human-in-the-loop AI done right is not a confirmation dialog. It is a policy-based approval architecture with risk tiers, enforcement gates, and full decision audit trails.
Synth as an MCP Server: Bridging Synthesis and Protocol
Synth runs as a native MCP server over JSON-RPC 2.0, exposing tool synthesis to Claude Code, Cursor, and any MCP-compatible client with built-in risk-based approval.
The Deloitte State of AI Report: What 3,235 Leaders Tell Us About 2026
Deloitte surveyed 3,235 leaders across 24 countries. 85% plan to customize AI agents, but only 21% have governance for them. 82% expect 10%+ of jobs fully automated within three years. The governance gap is the opportunity.
The Intelligence Slider: One Control to Balance AI Cost and Quality
One slider from 0 to 100 controls which AI model handles every request. Economy, Balanced, and Premium tiers with automatic failover — significant cost reduction while maintaining quality.
Building AI Agents with Our SDK
Build AI agents with our SDK: TypeScript-first, multi-provider, with streaming, built-in tools, MCP integration, and platform-level governance.
Kubernetes Hits 82% Production Adoption: The De Facto AI Operating System
CNCF 2025 survey: 82% of container users run Kubernetes in production. 66% of generative AI organizations use K8s for inference. Kubernetes is the de facto AI operating system.
2025 in Review: The Year Agentic AI Got Real
MCP exploded, OpenAI adopted it, Gartner predicted 40% agentic AI failure, and the EU AI Act began enforcement. A comprehensive look at the year agentic AI moved from hype to reality.
AI Supply Chain Security: A Comprehensive Guide for Enterprise Teams
Malicious package uploads are up 156% year over year per Sonatype. This guide covers every known AI supply chain attack vector and the defense layers enterprises need to stop them.
Getting Started with Synth: Your First Synthesized Tool in 5 Minutes
Install Synth, configure a provider, and synthesize your first AI tool in five minutes. A hands-on tutorial covering CLI commands, providers, capabilities, and the human approval flow.
MCP Joins the Linux Foundation: The Future of AI Integration Standards
Anthropic donated MCP to the Linux Foundation on December 9, 2025. Co-founded by Anthropic, Block, and OpenAI, the Agentic AI Foundation makes MCP an enterprise-safe open standard.
From Demo to Production: Why Most AI Agents Never Ship
Most AI agents work in Jupyter but fail in production. The demo-to-production gap is a category of problems: credentials, concurrency, monitoring, cost, audit. The platform eliminates the gap by design.
S&P Global: 42% of Companies Abandoned Most AI Initiatives in 2025
S&P Global found 42% of companies abandoned most AI initiatives in 2025, up from 17% the prior year. The average org scrapped 46% of POCs. Here are the root causes and what the successful 58% do differently.
MCP Turns One: The Protocol That Changed AI Integration
MCP launched one year ago and became the universal standard for AI tool integration. OpenAI adopted it in March 2025. Over 10,000 MCP servers exist. Here is what the first year revealed and what still needs solving.
NullifAI: How Malicious Models Bypass AI Security Scanners
ReversingLabs found malicious ML models on Hugging Face using NullifAI to bypass Picklescan. A new attack vector targeting model artifacts, not source code. Here is what enterprises need to know.
Multi-Cloud AI Without Lock-In: Practical Architecture Patterns
94% of IT leaders fear vendor lock-in. Here are the four layers of AI lock-in and the architecture patterns that prevent each one — with practical tests to verify portability.
Introducing Synth: On-Demand Agent Tool Synthesis
Synth eliminates pre-registered tool libraries. Describe intent in natural language, approve the synthesized tool, execute in a sandbox, destroy it. MIT licensed and available now via pip install synth-ai.
DLP and RBAC for AI Agents: Why MCP Servers Need Security That Frameworks Won't Give You
Why AI agents need DLP scanning and RBAC at the platform level — and why MCP servers and frameworks cannot provide it.
Why Gartner Named Agentic AI the #1 Strategic Technology Trend
Gartner named agentic AI the #1 strategic trend for 2025. 33% of enterprise software will include it by 2028 — but 40% of projects may fail. Here is what separates success from failure.
Enterprise AI Spending: Where IDC’s $632 Billion Goes
IDC forecasts $632B in AI spending by 2028, but McKinsey says only 6% capture real value. Here is where enterprise AI budgets actually go and where they should be redirected.
Credential Isolation for AI Agents: Architecture Patterns That Prevent Breaches
Most agent frameworks share all credentials across all tools. Credential isolation — scoped, time-limited tokens per tool and per task — is the architecture pattern that prevents breaches.
The Ultralytics YOLO Compromise: How GitHub Actions Became an Attack Vector
Ultralytics YOLO was compromised through GitHub Actions in December 2024. Source code was clean, malicious payload injected during build. CI/CD is the new attack surface.
AgenticLoops: Event-Driven Intelligence for Human-in-the-Loop Automation
Technical overview of AgenticLoops: event-driven behavioral triggers, identity resolution, and webhook dispatch for HITL automation.
Building AI Audit Trails That Satisfy SOC 2, HIPAA, and FedRAMP
SOC 2, HIPAA, and FedRAMP all require audit trails for automated systems. Here is how to build AI governance infrastructure that satisfies compliance requirements by design.
The GenAI Divide: Why 95% of Enterprise AI Pilots Fail to Scale
MIT found only 5% of AI initiatives produce measurable returns. The GenAI Divide is not about technology — it is about governance, ROI measurement, and the demo-to-production gap.
MCP Server Security: The Threat Model Nobody Is Talking About
MCP servers have no built-in sandboxing. They run with full host permissions. Here is the threat model enterprise teams must understand before deploying MCP in production.
Building an Immutable Audit System for AI Agents: Cryptographic Hashing and Compliance-Grade Logging
How the platform's Audit System records every agent action in tamper-evident, cryptographically hashed logs for enterprise compliance.
Gartner's Agentic AI Predictions: Separating Signal from Noise
Gartner named agentic AI the #1 strategic technology trend and predicts 40% of projects will be canceled by 2027. What these predictions mean for teams building AI systems today.
Extended Thinking: When AI Shows Its Work
Extended thinking lets AI models show their chain-of-thought reasoning before responding. Better answers for complex tasks, with full transparency into how the model reached its conclusion.
AI Agent Memory: Why Stateless Agents Fail in the Enterprise
Stateless AI agents waste tokens and frustrate users by forgetting everything between sessions. Here is how three-layer memory with session, user, and global context changes the equation.
Self-Hosting AI: From Cloud APIs to On-Premise Deployment
Why organizations are moving AI on-premise: data sovereignty, compliance, cost control, and latency. How the platform deploys on any Kubernetes cluster from EKS to air-gapped bare metal.
The McKinsey AI Adoption Gap: 88% Adopt, Only 6% Profit
McKinsey found 88% of organizations use AI, but only 6% are high performers capturing real value. The gap between adoption and impact is the defining challenge of enterprise AI.
Multi-Agent Orchestration: How the Platform Coordinates Specialized AI Agents
How the platform orchestrates ChatMode, CodeMode, Research Mode, and AgenticWorkflows under unified governance with shared audit trails.
Running AI Models Locally with Ollama: A Complete Guide
Run Llama, Mistral, Qwen, and dozens of other open-source LLMs locally with Ollama. Complete guide covering installation, model selection, API usage, and the platform's SmartModelRouter integration.
AI Governance Frameworks Compared: NIST, EU AI Act, and ISO 42001
NIST AI RMF, the EU AI Act, and ISO/IEC 42001 are the three frameworks shaping AI governance. Here is how they compare, where they overlap, and how the platform maps to each.
The PyTorch torchtriton Attack: What Every AI Team Needs to Know
The December 2022 PyTorch torchtriton attack exploited pip dependency confusion to steal SSH keys and system info. 2,717 downloads. Here is the full breakdown and how to prevent it.
Sandboxed Execution on Kubernetes: How the Platform Isolates AI Agent Tool Runs
Inside the platform's multi-layer sandbox for AI agent tool execution: code validation, process isolation, resource limits, and K8s containers.
Sandboxed Execution: Why Every AI Agent Tool Needs Isolation
Every AI agent tool should run in an isolated sandbox with explicit resource boundaries. Here is how the platform enforces Kubernetes-native tool isolation and why most agent frameworks get this wrong.
SmartModelRouter: How Automatic Model Selection Cuts AI Costs
SmartModelRouter analyzes task complexity and routes to the right model tier. Meaningful cost reduction while maintaining quality through intelligent automatic model selection.
Understanding Model Context Protocol: The Standard That Connects AI to Everything
Model Context Protocol (MCP) is the open standard connecting AI to external tools and data. Created at Anthropic and adopted by Claude, ChatGPT, Gemini, VS Code, and more. Here is how it works.
How We Built Credential Isolation: OBO Auth for AI Agent Tool Execution
Technical walkthrough of the platform's credential isolation: scoped tokens, OBO auth, and post-task cleanup for AI agent tool execution.
FedRAMP and AI: Navigating the Federal Authorization Landscape
Azure OpenAI, Google Gemini, and Anthropic Claude have all achieved FedRAMP High authorization through different cloud providers. Here is the current state of AI in the federal authorization landscape.
Integration Debt: The Hidden Cost Killing Enterprise AI Projects
Integration debt is the hidden cost killing enterprise AI. The tool registration model used by LangChain, CrewAI, and AutoGen creates linear cost growth. The platform offers a fundamentally different approach with MCP Workshop, Workflow Builder, and Synth tool synthesis.
Building SmartModelRouter: How We Route Across Multiple LLM Providers With Minimal Overhead
Technical deep-dive into SmartModelRouter: routing AI agent requests across 15+ model families through 5 provider integrations with automatic failover and cost optimization.
Building AI Agents That Actually Work in Production
S&P Global found 42% of companies abandoned most AI initiatives in 2025. The demo-to-production gap is real. Here are the production requirements that separate AI agents that ship from those that get shelved.
The EU AI Act: A Compliance Timeline for Technology Leaders
The EU AI Act entered into force August 1, 2024. Key deadlines span from 2025 to 2027. This timeline guide covers risk classifications, GPAI rules, penalties, and what US companies deploying AI in EU markets need to know.
80% of AI Projects Fail: What RAND Corporation's Research Actually Says
RAND Corporation found over 80% of AI projects fail, twice the rate of non-AI IT projects. The root causes are organizational, not technical. Here is what the research actually says and what to do about it.
Open Source AI in 2025: PyTorch, Hugging Face, and the New Stack
PyTorch leads model training at 63% adoption. Hugging Face is the model hub. vLLM and SGLang handle inference. This is the open source AI stack in 2025 and what it means for enterprise AI deployment.
Why We Started AgenticWork: Security-First AI Agents for the Enterprise
Why we built our platform: security-first, multi-model, multi-provider AI agent infrastructure for the enterprise.
The NIST AI Risk Management Framework: A Practical Guide for Engineering Teams
A practical engineering guide to NIST AI RMF 1.0 and AI 600-1 for generative AI. Learn how the four core functions map to real engineering decisions and how the platform implements each function.
Kubernetes for AI Workloads: Why K8s Became the AI Operating System
CNCF survey data shows 80% of container users run K8s in production and 54% run AI/ML on K8s. Kubernetes has become the operating system for AI workloads. Here is why.
What Is Agentic AI? A Technical Primer for Engineering Leaders
Gartner named agentic AI the #1 strategic tech trend for 2025. This technical primer explains the difference between chatbots, copilots, and agents, and introduces the Synth framework for production deployment.
The Case for Private AI Infrastructure
Data sovereignty, compliance, cost control, and latency all make the case for private AI infrastructure. Here is why self-hosted AI on Kubernetes is becoming the enterprise default.