Manufacturing IoT
AI Agents Could Become the Next Layer of IoT Platforms
Published by IOT San Jose Research & Editorial Team

In October 2025, AWS published a reference design for a device management agent. A user types a request in plain language, such as listing the devices in a home or changing a Wi-Fi network, and an agent picks the right tool, calls a function, and replies in a sentence. The architecture uses seven core tools, five database tables, and layers of authentication and guardrails. It is a small example, but it points to a large change: IoT platforms may soon be operated by conversation and by software agents, not only by dashboards and rules.
Every IoT platform today follows a similar pattern. Devices send data, the platform stores it, and people build dashboards, alerts, and rules on top. That model works, but it puts the burden of interpretation on humans, and fleets are now too large for people to watch. Agents offer a different division of labor. They can read the data, reason about it, and take approved actions. Whether that becomes the next layer of the platform depends on how well the industry handles trust.
What an Agent Adds to a Platform
A conventional rule says, "if temperature exceeds 80, send an alert." An agent can be given a goal instead, such as "keep this cold room within range and tell me only when intervention is needed." To meet it, the agent reads sensor streams, checks maintenance history, looks up a device's configuration, decides whether the problem is a failing sensor or a failing compressor, and either fixes the setting or opens a work order. The difference is not magic. It is the ability to combine several tools and several data sources to reach an outcome, with a language model deciding the order of steps.
SecuriThings launched an Agentic Device Orchestrator in March 2026 that lets administrators state a goal, for example checking remaining storage on all security cameras in a wing, and has agents build and run the workflow. The same agents handle vulnerability scanning, patch deployment, and password rotation across mixed device fleets. The company cites an IoT device management market of about 8.8 billion dollars in 2025 and the IoT Security Foundation's finding that 60 percent of IoT breaches are tied to outdated software. Those figures are the vendor's framing, but the underlying problem, a flood of unpatched devices, is well known.
Where the Agent Sits in the Stack
| Platform layer | Traditional approach | With an agent |
|---|---|---|
| Device connectivity | Protocols, certificates, onboarding workflows | Agent guides onboarding and diagnoses failed connections |
| Data and analytics | Stored time series, dashboards, scheduled reports | Agent answers questions in plain language and flags anomalies |
| Rules and automation | Fixed if-then rules written by engineers | Agent proposes or runs multi-step actions toward a goal |
| Fleet operations | Manual update campaigns and ticket queues | Agent plans rollouts, watches results, and rolls back on failure |
| Security | Periodic scans and alert triage | Agent investigates alerts and applies approved remediation |
In every row the agent does not replace the layer. It sits on top and uses the layer's functions as tools. That matters for platform buyers, because it means a good agent depends on a good underlying platform. A messy device registry or inconsistent data model will produce a confused agent, however capable the language model is.
The Trust Problem Comes First
Giving software the ability to act on physical equipment raises the stakes. A chatbot that gives a wrong answer wastes a minute. An agent that locks the wrong doors or shuts down the wrong chiller can cause real harm. Critics of orchestration platforms also point out that a central agent with authority over thousands of devices becomes an attractive target, because compromising it could hand an attacker control of the entire fleet.
The AWS design addresses this in the way experienced engineers would expect. Authentication happens before any request reaches the agent. Access is checked again at the gateway and inside each function. Each user session runs in its own isolated container. Guardrails help block prompt injection, the technique where hostile text tries to hijack the agent's instructions. Those layers are the template. Any agent that can act on devices needs the same principle: the model proposes, but a separate, deterministic layer decides what is permitted.
Three Levels of Autonomy
It helps to separate what an agent is allowed to do, because buyers often confuse the levels.
- Advise. The agent answers questions and suggests actions. A person approves everything. This is the safest level and delivers real value by saving time on diagnosis.
- Act with approval. The agent prepares a complete change, such as a firmware rollout to one building, and waits for a human to confirm. This suits most operational changes today.
- Act within limits. The agent executes pre-approved actions inside tight boundaries, for example restarting a stuck gateway but never touching a safety system. This level needs strong logging and a reliable way to undo changes.
Most organizations will spend a long time at the first two levels, and that is reasonable. The goal is to earn trust gradually, using audit trails to see what the agent would have done before letting it do it.
Why It Matters in the South Bay
The Bay Area is where both halves of this story are built. Cloud providers and platform companies design the agent frameworks, while device makers, facility operators, and manufacturers run the fleets. Companies such as JFrog, whose connect product is covered in our article on IoT device management from Sunnyvale, already handle the software supply chain that agents will rely on. The broader movement toward autonomy is discussed in our look at agentic AIoT.
A typical use case is a commercial property manager responsible for buildings across San Jose, Santa Clara, and Sunnyvale. This is an illustrative scenario. Today the team juggles separate systems for access control, cameras, HVAC, and meters. An agent connected to those systems could answer "which buildings have devices running firmware older than the security baseline" and prepare a staged update plan, while a human reviews and approves. The saving is not dramatic in any single task. It comes from removing the hours of cross-referencing that fill an operations week.
A Worked Example: A Stuck Firmware Rollout
Consider a facility group with 2,000 networked cameras and badge readers across a dozen South Bay buildings. This is an illustrative scenario. A security advisory requires a firmware update on one camera model within two weeks. Today an administrator would query the inventory, group devices by site, schedule maintenance windows, run the update in batches, and chase the failures by hand. It is easy work to describe and tedious to carry out.
With an agent in advise mode, the administrator asks which cameras run affected firmware and which buildings have the most of them. The agent returns a list, notes forty devices that have been offline for weeks, and proposes a three-wave plan that starts with a low-risk building. The administrator approves wave one. The agent runs it, watches for failures, and pauses when more than two percent of devices fail to report back. A human decides whether to continue, and the full sequence is logged.
The agent did not make a strategic decision. It did the cross-referencing, drafted the plan, and watched the rollout, which are the tasks that consume operations time. The value is real because it frees skilled people for the exceptions. The risk is also visible, because the same agent with unlimited authority could have pushed a bad update to all 2,000 devices at once. That is why limits, approvals, and rollback are part of the design, not optional extras.
Implementation Advice
- Fix the data model first. Clean device inventory, consistent naming, and accurate location data matter more than the choice of language model.
- Expose small, safe tools. Give the agent narrow functions such as "read device status" and "schedule update" instead of broad system access.
- Separate permission from reasoning. Enforce access rules in code outside the model, and log every call with the user and reason.
- Start in advise mode. Run the agent in a read-only role, compare its suggestions with what operators actually did, and widen its authority only after the match is strong.
- Test hostile inputs. Try prompt injection through device names, log messages, and ticket text, since agents read text that attackers may influence.
Common Challenges
| Challenge | Why it matters |
|---|---|
| Prompt injection | Text inside device data can attempt to override the agent's instructions |
| Single point of compromise | A powerful orchestrator is an attractive target and needs strict isolation |
| Unclear accountability | When an agent acts, teams need clear records of who authorized what |
| Cost and latency | Language model calls add cost, so frequent routine checks may be cheaper as plain rules |
| Platform lock-in | Agent frameworks tied to one cloud can make later migration harder |
What Comes Next
Researchers are already writing about an internet of agents, where devices expose capabilities that agents discover and use. Espressif announced an agents platform for its chips in late 2025, and Microsoft has described physical AI for industrial operations through its IoT tooling. The direction is consistent. Platforms will become the control plane that agents use, and the differentiator will be which platform makes agents safest to deploy. For buyers, the sensible stance is curiosity with discipline: pilot agents where the stakes are low, demand strong audit trails, and keep humans in the loop until the evidence says otherwise.
