Blog Deep dives on LLMs, AI agents, and tool calling
Closing the Loop: A Self-Evolving Defense
How we built an autonomous loop that discovers new prompt injection techniques, scores them against real models and Defender, and files a PR back to the package
Hiskias Dingeto · 
Autoresearch-Charged Action Search and Benchmark for AI Agents
How we fine-tuned a 109M embedding model for tool retrieval that beats 8B models on ToolRet, MetaTool, and ToolBench. Methodology and the autonomous loop.
Will Leeney · 
From Handshake to Integration: Your Complete Guide to Technology Partnerships
Learn how to choose the right partnership model, negotiate effectively, launch with impact, and build integrations that keep customers happy — with practical advice for the agentic AI era.
AI Agent Failures Are a Permissions Problem
The largest public dataset of AI agent misbehaviour ranks overeagerness first at 43.4%. The most common failure is an agent doing more than it was asked.
Guillaume Lebedel · 
EU AI Act Deadline Moved to 2027. Agent Access Didn't.
The EU AI Act deadline for high-risk AI moved to December 2027. AI agent access controls and audit trails did not change. What to fix before the new date.
Guillaume Lebedel · 
Slack MCP Server: Capabilities, Limitations, and Alternatives
A review of the official Slack MCP server: capabilities, limits, and an alternative, to help you choose which to use for your AI agent.
Jira webhooks for AI agents: turn issue and sprint events into cross-app agentic workflows
Turn Jira events into cross-app agentic workflows, cut token spend and lag, and get a full audit trail of every action.
Multi-Agent Systems Coordinate. They Don't Execute.
Multi-agent systems fail in production 41-87% of the time, and most breakdowns land at the handoff into your systems of record. Coordination isn't execution.
Guillaume Lebedel · 
Salesforce webhooks for AI agents: turn CRM events into cross-app agentic workflows
Turn Salesforce events into cross-app agentic workflows, cut token spend and lag, and get a full audit trail of every action.
ServiceNow webhooks for AI agents: turn IT service events into cross-app agentic workflows
Turn ServiceNow events into cross-app agentic workflows, cut token spend and lag, and get a full audit trail of every action.
Webhooks for AI agents: why agentic workflows need them purpose-built
AI agents can act the moment a system changes, using webhooks. But webhooks built for software hit three limits in agentic workflows. Here is how to fix each.
The New MCP Spec Won't Get Your Agents Into Production
The 2026-07-28 MCP spec hardens transport, auth, and lifecycle. It still won't decide which actions your agent may run, who approved them, or what's logged.
Guillaume Lebedel · 
The EU AI Act Will Grade Your Logs, Not Your Model
On August 2, the EU AI Act's high-risk rules take effect. Article 26 makes AI agent deployers keep six months of per-action logs. It's an action-layer job.
Guillaume Lebedel · 
Why a Free Model Won't Make Your AI Agent Cheaper
GLM-5.2 shipped free under MIT and tops the open-weight agent benchmarks. So why can a free model raise your agent's token bill? The cost is tools, not license.
Guillaume Lebedel · 
Inside StackOne's Agent: Built on A2A, ADK, and Gemini
StackOne's agent runs on Google's open agent stack: the A2A protocol, ADK, and Gemini. Connect StackOne once and reach 300+ integrations from any A2A agent.
Guillaume Lebedel · Closing the Loop: A Self-Evolving Defense
How we built an autonomous loop that discovers new prompt injection techniques, scores them against real models and Defender, and files a PR back to the package
Hiskias Dingeto · 
Salesforce's $3.6B Fin Deal: Why the Moat Is Integration
Salesforce is paying $3.6B for Fin, an AI agent with its own model. The real prize is integration and distribution, not the model. Here's why it matters.
Guillaume Lebedel · 
Why Human-in-the-Loop Fails as an AI Agent Guardrail
An AI agent asked permission before running up a $6,531 AWS bill, and the human approved it. Why guardrails belong in credentials, not approval prompts.
Guillaume Lebedel · 
How I Built an AI Hiring Analyst in Under an Hour
A Head of People connected Claude to Ashby through StackOne's MCP gateway to answer hiring questions from real-time data via chat. No code required.
AI Agent Identity: Lessons from Fedora's Rogue Agent
An AI agent ran inside Fedora for two weeks under a compromised contributor account and got code merged. Why it's an identity failure, not a model failure.
Guillaume Lebedel · 
The Best MCP Gateways for Dust in 2026
Best MCP gateways for Dust: keep Dust's tool stakes and admin controls, add StackOne for Workday, ServiceNow and SAP actions. Composio and Zapier compared.
Romain Sestier · 
The Best MCP Gateways for Lyzr in 2026
The best MCP gateways for Lyzr Agent Studio, compared — StackOne, Composio, Workato, Zapier, Arcade — and the gaps a gateway fills. Updated June 2026.
Romain Sestier · 
The Best MCP Gateways for Retool in 2026
Best MCP gateways for Retool Agents: StackOne, Composio, Zapier, Merge, Workato compared. Retool caps 20 MCP servers per agent; audit logs skip tool calls.
Romain Sestier · 
The Best MCP Gateways for Salesforce Agentforce in 2026
The best MCP gateways for Salesforce Agentforce, compared: StackOne, MuleSoft, Zapier, Composio, Merge — under Salesforce's ~20-tool budget. June 2026.
Romain Sestier · 
The Best MCP Gateways for Claude Code in 2026
The best MCP gateways for Claude Code: StackOne vs Docker vs Composio vs Arcade — one pinnable managed-mcp.json URL, per-user OAuth, audit logs.
Romain Sestier · 
The Best MCP Gateways for ChatGPT Enterprise in 2026
The best MCP gateways for ChatGPT Enterprise in 2026: StackOne, Zapier, Composio, Pipedream, and Merge — now that read-write MCP apps are in beta.
Romain Sestier · 
The Best MCP Gateways for Claude in 2026
The best MCP gateways for Claude: StackOne, Composio, Zapier, Arcade, Merge, and Docker compared — Claude's audit logs don't capture connector activity.
Romain Sestier · 
The Best MCP Gateways for Copilot Studio in 2026
The best MCP gateways for Copilot Studio, compared: StackOne, Zapier, Workato, and Microsoft's native stack — wizard auth, the 128-tool cap, DLP and ACP.
Romain Sestier · 
The Best MCP Gateways for Gemini Enterprise in 2026
The best MCP gateways for Gemini Enterprise (formerly Agentspace), compared: StackOne, Composio, Zapier, Arcade vs Google's native connectors.
Romain Sestier · 
The Best MCP Gateways for Glean in 2026
Best MCP gateways for Glean: keep Glean's native governance, add StackOne for Workday, SAP, UKG and ServiceNow actions. Composio and Zapier compared.
Romain Sestier · 
The Best MCP Gateways for Microsoft Copilot in 2026
The best MCP gateways for Microsoft Copilot in 2026: StackOne, Zapier, and Workato compared on the 128-tool cap, DLP, and Agent 365 — M365 Copilot.
Romain Sestier · 
The Best MCP Gateways for n8n in 2026
Best MCP gateways for n8n: StackOne, Zapier, Composio, Pipedream compared. n8n has native MCP nodes but no MCP allowlist; audit stops at the workflow edge.
Romain Sestier · 
The Best MCP Gateways for the OpenAI Agents SDK in 2026
The best MCP gateways for the OpenAI Agents SDK, compared — StackOne, Composio, Pipedream, Arcade — as Agent Builder winds down. Updated June 2026.
Romain Sestier · 
The Best MCP Gateways in 2026, Compared
We compared 12 MCP gateways — StackOne, Zapier, Composio, Microsoft, Docker and more — on auth, multi-tenancy, governance, and pricing. Updated June 2026.
Romain Sestier · 
OAuth for AI Agents: How It Works and the Benefits of Owning the OAuth App
OAuth for AI agents, explained: how applications and authorizations work for agentic execution in production, and why owning your own OAuth app matters.
Autoresearch-Charged Action Search and Benchmark for AI Agents
How we fine-tuned a 109M embedding model for tool retrieval that beats 8B models on ToolRet, MetaTool, and ToolBench. Methodology and the autonomous loop.
Will Leeney · 
Detection Isn't the Hard Part
How we cut Defender's false positive rate by ~97% on real connector data with two fixes: distribution-matched hard negatives and structural field filtering.
Hiskias Dingeto · 
Training a 22MB Prompt Injection Classifier
How we built and iterated on the 22MB ONNX classifier inside Defender: backbone selection, training data, freeze-layer optimization, and the inference pipeline.
Hiskias Dingeto · 
How AI Is Changing the True Cost of SaaS Integrations
I managed 700 SaaS connectors at Tray.ai. Over 400 were unused. AI is rewriting integration cost economics at every step — here's what changes.
RAG vs MCP for AI agents: When to index, when to fetch live
How RAG and MCP feed AI agents in production. When to index, when to fetch live, and when production agents need both.
How to Build an AI Marketing Analytics Agent with Claude + an MCP Gateway
Connect Claude to GA4 and HubSpot through StackOne's MCP gateway. Surface bot traffic, funnel gaps, and signup attribution your dashboard doesn't show.
How I Vibe-Coded a Customer Success Hub Using StackOne's AI Native Integration Platform
How a non-developer vibe-coded a customer success dashboard pulling HubSpot, Pylon, PostHog, Databricks, and Fireflies through StackOne's MCP gateway.
Supercharging Research with Agents
Building an end-to-end research system live at a talk, from context files to a second brain with knowledge graphs.
Will Leeney · 
MCP Token Optimization: 4 Approaches to Fix Context Window Bloat
MCP tools eat context windows two ways: schema bloat and response bloat. Four approaches to fix it, with sourced benchmarks and honest tradeoffs.
Guillaume Lebedel · 
The Two-Loop Architecture: How AI Agents Use CLI and MCP Together
MCP vs CLI is a false binary. Claude Code already uses both. Here's the two-loop architecture — CLI for local tools, MCP for remote SaaS — with setup examples.
Guillaume Lebedel · 
MCP vs CLI for AI Agents: When Each One Wins
MCP vs CLI for AI agents — CLI wins for local dev tools, but there's no Workday CLI. A system boundary framework for choosing the right interface.
Guillaume Lebedel · 
MCP vs API: What 200+ Connector Builds Taught Us
MCP wraps APIs, it doesn't replace them. After building 200+ connectors that serve both, here's when each approach wins.
AI Agent Testing: What We Found Testing Thousands of Agent Tools
We tested thousands of API actions across dozens of SaaS platforms to see if AI agents could use them. Most failures looked like successes — 200 status codes hiding errors, schemas missing fields, agents skipping steps.
How I Built a Customer Support Dashboard in 2 Days Using StackOne MCP
How a solutions engineer used Claude Code and StackOne's MCP tools to build an internal support triage dashboard in 2 days. Includes automated prioritisation, investigation prompts, and response drafting.
Max Strivens · From Handshake to Integration: Your Complete Guide to Technology Partnerships
Learn how to choose the right partnership model, negotiate effectively, launch with impact, and build integrations that keep customers happy — with practical advice for the agentic AI era.
MCP vs A2A: Architecture, Security, and When to Use Each
MCP vs A2A: what each protocol standardizes, how they differ, their shared security risks including indirect prompt injection, and when to use one, both, or a hybrid architecture.
Gartner Says 60% of AI Agent Deployments Will Fail — The Reason Is Integration
Gartner predicts 60% of early agentic AI implementations will fail due to integration gaps. We break down their latest research and show how StackOne's connectivity infrastructure addresses every key recommendation.
Richard South · 
How AI Consultancies Can Get Their Clients' Agents Out of Pilot Purgatory
95% of enterprise AI pilots fail to reach production. The bottleneck isn't models or prompts — it's connectivity. Here's the three-pillar framework consultancies need to deliver agents that actually work.
Richard South · 
We Rebuilt Our Marketing Site with Astro, Claude Code, and Cloudflare. In a couple of evenings.
We replaced Webflow with Astro 5 + Cloudflare Pages in two weeks using Claude Code as the primary builder. 42 pages, 50 PRs — non-engineers ship independently.
Guillaume Lebedel · 
Indirect Prompt Injection in MCP Tools: 10 Real Examples & Defenses
See how MCP injection (indirect prompt injection through MCP tool calls) threatens your AI agents across Gmail, Slack, Salesforce and 7 other integrations, with defenses.
Why AI Agents Are Changing the Integration Paradigm for B2B SaaS Platforms
Traditional 'build to us' integration strategies are becoming obsolete. Companies that don't enable their Agents to connect outward risk becoming legacy data sources.
Richard South · 
MCP Code Mode: Keeping Tool Responses Out of Agent Context
Anthropic's code_execution processes data already in context. Custom MCP code mode keeps raw tool responses in a sandbox. 14K tokens vs 500.
Guillaume Lebedel · 
Comparing BM25, TF-IDF, and Hybrid Search for MCP Tool Discovery
Benchmarking BM25, TF-IDF, and hybrid search for MCP tool discovery across 916 tools. The 80/20 TF-IDF/BM25 hybrid hits 21% Top-1 accuracy in under 1ms.
Guillaume Lebedel · 
Indirect Prompt Injection Defense for MCP Tools: A Technical Guide
MCP tools that read emails, CRM records, and tickets are indirect prompt injection vectors. Here's how we built a two-tier defense that scans tool results in ~11ms.
Guillaume Lebedel · 
The Trillion-Dollar Wake-Up Call: Make Your Agents More Capable or Become a Legacy Platform
Over $1 trillion in SaaS market value evaporated in a week. The market is pricing in a structural shift: SaaS platforms that can't power connected agents will become backend data sources.
Richard South · 
Unified API Limitations for AI Agent Integration: 7 Ways They Break
Unified APIs work for traditional software but fail AI agents in specific, measurable ways. Here are 7 problems we found after building both approaches, with examples from Workday, Greenhouse, Jira, and Salesforce.
Guillaume Lebedel · ![120+ Agentic AI Tools Mapped Across 11 Categories [2026]](/_astro/ai-agent-tools-landscape-2026.D9-zOF-k_VatgH.png)
120+ Agentic AI Tools Mapped Across 11 Categories [2026]
Map of 120+ agentic AI tools across 11 categories — frameworks, no-code builders, observability, coding agents, enterprise platforms, and more. Updated Q1 2026.
Romain Sestier · 
AI Agents Meet Gamma: Automating Presentations with StackOne Connectors
Two real examples of how we use Gamma with AI Agents at StackOne -- from prepping sales decks to presenting company structure -- all powered by StackOne connectors.
Richard South · 
Building Semantic Search with Enhanced Embeddings
Adding semantic search to 10,000+ connector actions using synonym-enriched embeddings and Turbopuffer. Benchmarks, architecture, and results.
Will Leeney · 
Beyond the Lab: Solving the Context Gap to Scale Enterprise AI Agents
Most AI agents are stuck in the lab phase. The missing piece isn't intelligence — it's context. Here's how to bridge the gap between pilot and production.
Richard South · 
Your Agent Toolkit: MCP Server, SDK-based Toolset, or Both?
MCP promises standardization but framework support varies wildly. SDKs give control but mean vendor lock-in. The hybrid approach fills gaps that pure options leave open.
Guillaume Lebedel · 
Automating Meeting Sharing with a Custom Fireflies MCP Server
StackOne's Fireflies MCP connector adds write operations the Fireflies MCP server lacks. Automate meeting sharing with Claude Code.
Romain Sestier · 
Agentic Context Engineering: Why AI Agents Kill Their Own Context Windows
AI agents exceed their context windows without knowing it. Six failure patterns and seven survival architectures for agentic context engineering.
Guillaume Lebedel · 
Notion MCP Server: Capabilities, Limitations, and Alternatives
The official Notion MCP server handles search and retrieval well. This post covers what it can't do and when the StackOne MCP server is the right alternative.
Romain Sestier · 
MCP: What's Working, What's Broken, and What Comes Next (what a year in production has taught us)
What a year in production has taught us about the Model Context Protocol -- from authentication gaps to Code Mode and the Linux Foundation transfer.
Guillaume Lebedel · 
Building an Agentic Researcher
How StackOne built a self-improving agentic researcher that continuously discovers AI agent integration use-cases online and refines its search queries over time using Slack reaction feedback and DSPy optimization.
Will Leeney · 
How to Integrate with Workday Learning via SOAP API
If you're looking to build a Workday Learning integration to push content and completions to Workday you will need to use the SOAP API, this guide covers the basics - from authentication and permissions to managing learning content and completions sync.
Max Strivens · 
When Your AI Agent Needs Contractors, Not Apprentices: Building Async Subagents
Explores the difference between using apprentice-like tools versus contractor-style independent agents for handling specialized tasks.
Nicolas Belissent · StackOne named a Gartner Cool Vendor in HR Technology 2025
StackOne has been recognized by Gartner as a Cool Vendor in HR Technology 2025 for innovative products that strengthen the employee experience.
Romain Sestier · 
Building an Eval-Driven AI Feature at StackOne
Integrations break. Our new AI agent translates cryptic provider errors into clear fixes. How we built it -- and why starting with evals on day one made all the difference.
Will Leeney · 
Bring AI Agents to Work with StackOne
How StackOne is solving the challenge of connecting AI agents to perform actions across complex enterprise software stacks.
Matt Carey · 
Announcing StackOne Documents API - One API and File Picker for Your Customers' Knowledge
Integrate once to access every customer document - across cloud storage and knowledge bases - with a native picker and real-time, privacy-first architecture.
Romain Sestier · 
Rethinking MCP: From App to Workflow Servers
After spending months building and using MCP servers for daily tasks, exploring decoupling tools from traditional app-based servers to allow bundling tools into workflow-specific servers.
Bryce Tozer · My Dream AI Client
After months of deeply using different AI clients, identifying key product features that would make the ideal AI client.
Bryce Tozer · A New Direction for Tool Orchestration
Discover how Mastra gave us the control we needed to scale API integrations with AI.
Nicolas Belissent · 
Announcing our $20M Series A Led by GV (Google Ventures) to Power the Universal Integration Layer
$20 million Series A funding round led by Google Ventures, with participation from Workday Ventures and XTX Ventures, to continue building a Universal Integration Layer for B2B SaaS and Agents.
Romain Sestier · 
Tools not Rules
How MCP is helping us build a more adaptable agent.
Nicolas Belissent · 
Automating tasks with MCP
How StackOne uses MCP servers and tools to automate internal tasks such as raising Jira tickets.
Bryce Tozer · 
StackOne ToolSet: Enabling Agentic Integrations for Hundreds of Enterprise SaaS
Introducing the StackOne ToolSet, an open-source, experimental SDK tailor-made for AI agent development.
Matt Carey · 
Announcing StackChat - the most advanced API Assistant
Announcing StackChat: the most advanced API knowledge agent.
Guillaume Lebedel · 
Announcing StackOne's IAM API: Retrieve Comprehensive Data on User Access Across More than 80 SaaS Applications
Unveiling the latest IAM API, allowing deeper integration into customers' HR tech stack with access data across 80+ SaaS applications.
Romain Sestier · 
Navigating The Shift From Features To Outcomes In SaaS
There's been a significant shift in buyer priorities, moving from features and functions, to outcomes.
Romain Sestier · 
Why Privacy And Real-Time Are Non-Negotiable When Integrating with HRIS
Exploring the key reasons why real-time provisioning and offboarding is crucial for HRIS integration.
Romain Sestier ·