If you haven’t come across OpenClaw yet, here’s the short version: it’s an open-source personal AI assistant that runs on your own machine. Mac, Windows, or Linux. You talk to it through whatever chat app you already use (WhatsApp, Telegram, Discord, Slack, Signal, even iMessage), and it can actually do things. It has access to your file system, can browse the web, run scripts, manage your email, calendar, and much more through a growing ecosystem of skills and plugins. It has persistent memory, so it remembers your preferences and context across conversations. Think of it as a 24/7 AI teammate that lives on your computer and gets better the more you use it.
Like everyone else in the AI space, we jumped into OpenClaw testing early on. However, true to our core company philosophy, we didn’t just want to play around with it. We wanted to find use cases that deliver real value as soon as possible.
We didn’t have to look far. Every one of us is always short on time, and we all need an extra pair of hands to get everything done. So we decided to put the agent to the ultimate test: could OpenClaw work as a genuine Executive Assistant?
Turns out, with the right setup, it absolutely can. With a big asterisk: none of this is perfect yet, and we’ll be upfront about that throughout this series.
Why an Executive Assistant?
The EA role is one of the most demanding jobs out there. It requires managing communications across dozens of channels, keeping calendars in order, tracking promises and follow-ups, monitoring relevant news, and doing it all with the right tone depending on who you’re talking to. It’s also a role where AI’s strengths (processing large amounts of information, endless patience, and being available around the clock) could genuinely shine.
The challenge was turning a raw AI agent into something that could actually handle all of this reliably. That meant building plugins, integrations, security layers, and an entire ecosystem of tools around it.
What We Built
The project grew into something far more comprehensive than we initially expected. Before we go through the individual capabilities, it’s worth understanding how this actually works under the hood.
There’s a primary agent that handles most of the communication and big-picture management. Think of it as the main brain. But it doesn’t work alone. Each plugin we built comes with its own set of specialized sub-agents: dedicated experts that handle specific tasks within their domain. The M365 plugin, for example, has a sub-agent that’s an expert on composing emails (it knows our communication etiquette inside out), another that processes and organizes incoming messages, and more. The Micro Apps plugin has a builder agent that constructs apps using our framework, a project manager that handles larger multi-step tasks, execution agents that the project manager can spin up as needed, a researcher agent that knows all the research tools available, and several others.
This architecture isn’t just for show. It solves real problems. Each sub-agent carries only the context it needs, which helps with LLM context limitations. A sub-agent that only needs to know about email composition doesn’t need to be loaded up with health monitoring rules or micro app schemas. It also means each sub-agent can be deeply specialized, carrying detailed knowledge about its narrow domain without diluting the primary agent’s ability to coordinate everything.
The real value, though, isn’t in any single agent or capability. It’s in how the primary agent combines everything on the fly based on whatever the situation requires. It already has broad AI capabilities out of the box, and each plugin, sub-agent, or integration we add becomes another tool it can reach for. When you ask the EA for help, it doesn’t think in terms of “which feature do I use.” It pulls from its calendar access, email, communication skills, health data, micro apps, and whatever else is relevant, all at once, assembling a response or action that fits the specific moment. The magic is in the combination, not the parts.
Building this level of functionality wouldn’t be realistically possible without this layered approach. Some tasks need a dedicated expert agent. Others need teams of sub-agents working together. And the primary agent needs to know when to handle something itself and when to delegate. Getting that orchestration right has been one of the most interesting challenges of this project.
Here’s the big picture of what we’ve built so far.
Microsoft 365 Integration. Every EA needs email and calendar access. We built a proper M365 plugin that gives OpenClaw its own email, calendar, tasks, and contacts, plus carefully scoped access to shared accounts and resources. It monitors my inbox, group mailboxes like info@, and knows the purpose and context of each shared account. We’ll cover this in detail in a dedicated post.
Communication Intelligence. The agent doesn’t just read and write emails. It knows how to communicate. We built a communication etiquette system that adjusts tone depending on the situation. Talking to a coworker? Professional but casual. The CEO’s family member reaching out? Warm and personal. Group conversation with mixed participants? It reads the room and adapts. If the tone of the conversation shifts or new people join, it adjusts accordingly.
Organizational Memory. One of the weak points of current AI is context loss. To handle this, we built clear processes for how the EA organizes its mailboxes and calendar. Everything is categorized by status (processed, in-process, needs processing) so even if context is lost, the agent can immediately pick up where it left off.
Security Firewall. An AI that communicates with the outside world is a target for prompt injection attacks. This is a real concern, and the security of LLMs isn’t fully solved yet. So we teamed up with heySec and built a firewall that analyzes and filters all external content before it reaches the agent. Today it’s fully functional. More details coming soon on heySec’s channels.
Micro Apps. Sometimes the help you need from an EA isn’t about emails or chat. Sometimes you need it to help you organize everything that’s happening around you: ongoing projects, things that are constantly changing, information coming in from all directions, actions that need tracking. The world around an executive is chaotic, and keeping all of that structured and accessible is a job in itself. The hard part is that these needs change constantly, sometimes daily, and there’s no time to wait on development. So we built a plugin with a companion web and mobile app that lets you create, use, and change apps on the fly. We call it Micro Apps. More on this in a separate post.
Health Monitoring. Executive jobs are tough on your health. It’s easy to skip meals during busy times, ignore rising stress levels, or push through when your body is telling you to slow down. We built a companion app that syncs Apple Health data to a central database, giving the EA access to vitals, sleep patterns, eating habits, and workout data. It doesn’t just report. It proactively monitors trends and makes recommendations when they’re actually useful. This deserves its own deep-dive post.
Promise Detection. The EA monitors all of my communications and detects when commitments are made, like agreeing to meet someone at a specific time. If there’s no matching calendar event, the EA catches it and creates one automatically.
Morning Briefings. Every morning, the EA scans relevant news sources and social media channels and sends a briefing about everything the executive should know. The big unlock with LLMs is that there’s no language barrier. It monitors sources in Chinese, Estonian, Ukrainian, and more, surfacing relevant information regardless of the original language.
Not All Rainbows and Unicorns
I’d be lying if I said this was smooth sailing. We’ve had many failures along the way. We even had to completely restart and reinstall everything from scratch once, when the agent broke itself so thoroughly that fixing it wasn’t worth the effort.
And let’s be honest about where things stand right now: none of these capabilities are perfect. They’re more like MVP status. Things work as expected roughly 70 to 80 percent of the time. The other 20 to 30 percent? Sometimes that’s down to the real limitations of current LLM technology. Other times, it’s simply the reality of a project that’s still in its early stages. This is roughly a one-month-old system, and there’s a lot of maturing left to do on our side too. The agent forgets things it agreed to do. It silently fails on tasks and doesn’t tell anyone. It does the same task differently on Tuesday than it did on Monday, for no apparent reason. If you’re imagining a flawless digital employee, that’s not what this is. Not yet.
Context management has been one of the biggest challenges. Figuring out how to give the agent all the information it needs without drowning it in massive prompts. And yes, we once accidentally blew up the context window so badly that it burned roughly €200 a day in tokens just trying to process basic functions before immediately crashing.
We’ll cover the challenges and lessons learned in a dedicated post, because there’s a lot to share there.
Where We Are Now
There is a lot we have done already, and yet we have not even scratched the surface of what we have planned and what is possible. The EA is evolving every day, and with each iteration it becomes more capable, more reliable, and more genuinely useful. But it’s important to go into this with the right expectations. This is not a finished product. It’s a very promising work in progress, built on technology that has real and fundamental limitations today.
Over the coming weeks, we’ll publish detailed posts covering the M365 integration, Micro Apps, health monitoring capabilities, and the challenges we’ve faced along the way. Stay tuned, this is just the beginning.
Share this article
