In October 2026, I think we started to see the results of AI affecting everybody in the tech world, but not exactly in the direction they promised. Technically, engineering productivity has significantly grown – this is undoubted – but with a notable difference between different teams and projects. Some teams adapted their collaborative workflow to AI, some left it at the per-developer level, and some are still struggling with their ecosystem. Everything depends on so many factors.
But at the same time, a lot of individuals now have such a highly capable superpower that anybody without it simply cannot compete. For example, look at this CVE for PHP. Once released, it created a hole in millions of PHP-powered websites and computer systems. Or look at another CVE that affects any Linux server listening on TCP sockets.
We run our infrastructure on Debian Linux servers. We also have PHP-powered https://flespi.com and https://forum.flespi.com websites that are located outside our datacenter but are still resources maintained by us and under our responsibility.
So instead of moving forward, we and many other services on the Internet deployed upgrades to our servers. Day by day, one CVE after another… And it seems like the situation will get even worse soon. Too much code was produced in the past, and with these AI agents, it is much easier to research vulnerabilities on a mass scale. Hopefully, one day we will close all the gaps, but for now, prepare to deploy operating system, module, and server upgrades on a daily or weekly basis. And this is exactly the cost we pay for AI.
And have you ever thought about GPS tracking devices or IoT routers that use the same stack? They are open to the world as well and can actually be targets of hacking attacks. Maybe it doesn't make that much sense, but still…
Anyway, besides infrastructure upgrades, we issued a number of software updates, delivered new features, improved the consistency of our services, and were able to maintain 100% monthly uptime in October.
We integrated a new Hikvision protocol with video functionality.
Two new connectivity providers were supported in flespi to be used as a modem: emnify and onomondo.
Especially for our users in Ukraine, Poland, and the Middle East, to protect devices from electronic warfare, we enabled plugins to fix the message parameters before registration. Mostly, this is to address long coordinate jumps and future time reporting.
We also worked a lot on API consistency and documentation coverage. One of the results of this work is the standardization of API call responses across the platform.
For tachographs, we supported remote management of the TBA application. For certain protocols, we enabled caching of partially uploaded DDD files so that the device can continue sending only the missing file parts. A lot of effort also went into supporting Lisle racks, so that tacho card management will be seamless at a large scale.
The CDN storage system also got some attention. Initially, this was a simple file storage mechanism where you could upload a file, get a unique, unguessable link, and use it wherever needed. However, during the implementation of AI agents, when we gave them the possibility to use CDNs to publish temporary files and share links with users, we soon discovered that they not only wanted to share links but were also capable enough to build and publish real applications. So we started to adapt the underlying CDN engine for this purpose and have so far released a few important enhancements:
Possibility to replace a file by UUID – to modify the content of a previously published file, with a cache-control attribute included;
CDNs got their own domain name and path-level file addressing – basically, it means that a CDN can be treated as a website: you can upload content there and reference it by the path.
And right now, we are developing an alias property for CDNs, where you will be able to specify your own domain name for the website and address files there. For us, this is an important feature, as it will allow applications built by agents to be hosted under your own domain name in our infrastructure for simplicity, although agents are perfectly capable of maintaining files in third-party infrastructure as well.
We also worked a lot on the Wirepas protocol in order to provide a positioning engine, support various compatible devices for mesh networks, and, in general, make the indoor tracking system in flespi as powerful and flexible as the one we have for normal GPS-based tracking devices.
And finally, a lot of time was invested in flespi AI agents. Initially, we thought that they would be great autonomous workers for the telematics businesses of our users, assisting their employees with daily tasks. And they are perfectly good at it, able to reconcile a lot of information and carefully operate fleets with tens of thousands of devices.
So in October we were focused on enhancing their capabilities and connected flespi AI agents to the next systems:
Traccar - open-source GPS tracking software popular among flespi users;
Emnify and Onomondo - popular connectivity providers among our users (1NCE is migrating their API, so it is also in our plans; however, we need access to the v2 API for this);
We also implemented a generic HTTP connector so that a flespi AI agent can be connected to any 3rd party HTTP API (e.g., your own API or ERP) if provided with specifications.
Another interesting connector is Persona. It allows our users to set up a behavioural framework for agents and rebrand them as their own AI service. Paired with Agentbox, which provides an embeddable frontend component, this simplifies the provision of AI agent services under your own name.
As more and more users started working with our agents (we have 100% monthly growth in the number of agents and even higher growth in their work, measured across our user base), we discovered that all-or-nothing approval management (ask or allow) is quite difficult to manage. So we implemented granular approvals, allowing agents to work on certain components, e.g., the Telematics Hub, in auto mode, while platform entries are still subject to human-in-the-loop approval.
But around mid-October, we noticed that our agents are not only good at working with telematics but also have a great capability to set up computational automation (e.g., running without AI) and build applications for users.
So we started to adapt our harness to run a user-level automation system powered by code produced by our AI agents. This is exactly the low-level, top-level automation system our users often need in some tricky, rare cases, and now agents are capable of creating and maintaining it. I called it dark automation, and in the article, you can read how the agent is managing a smart tachograph scheduler in the background. Such an automation system – for some, it's about device firmware management; for others, it's about configurations, device glitches, health monitoring, and so on – simply works for different flespi users, addressing their particular needs.
Another distinct feature popped up from nowhere when we discovered that agents are pretty capable of writing web applications. All in all, they were equipped with the right (e.g., flespi) engineering mindset in their system instructions, grounded their scripts in API docs, and had a great training base. What we needed was just to equip them with a corresponding set of tools and configure the infrastructure – in other words, to adapt the harness for this magical brain.
And we did it. Now you can officially build cool applications with a flespi AI agent in just a few reasoning steps. Depending on the requirements, expect a cost of around 5–30 EUR per application. Applications are not specifically designed for flespi – in 10 minutes, I built a cool Teltonika FOTA dashboard to get a clear view of devices needing my attention, as well as an Emnify/Onomondo connectivity dashboard visualizing SIM statuses. For Wialon Hosting, it can even be a completely separate application based purely on Wialon data.
So far, apart from dashboards (which are very cool, by the way), I’ve tried a couple of 3D applications – flight data visualization for an aero club (an amazing one!) and a 3D indoor asset tracking portal I was shown. All these are great applications that a team of developers would build for you in a month. Offloading all the engineering and maintenance work to a flespi AI agent costs significantly less, with hosting and all the bells and whistles included.
And finally, one of the applications I developed with agents has been officially added to our KB. It is an n8n-style engine to build device automations. It is built with AI, but the automations it creates do not use AI when running (they utilize available flespi entities such as plugins, calculators, webhooks, etc.). It closes the gap we had in making various device-level automations easy to construct and maintain. Please read the full article and try this application on your flespi account. The application itself and its documentation were fully implemented by our agent, and I think it is a great showcase of how capable our agents are.
Let me remind you that each flespi account has a number of free AI credits available per month, so if you haven’t tried the AI agent yet, I recommend spending them on developing an app you’ve always dreamed about but never had enough time or skills to implement. Now is exactly the time!