On the evening of October 31, DeepSeek released the Super-agent architecture. Its core idea is to use a dual-layer architecture, event-driven design, plus an external Watchdog, letting the system's underlying layer take over session management and context synchronization—in other words, AI begins to direct AI to do the work, and even tries to reproduce and fix problems when issues arise within itself. In the same week, the National Bureau of Statistics announced that the October manufacturing PMI fell to 49.3%, with both medium-sized and small enterprises having fallen below the boom-or-bust line. On one side, agents are increasingly able to "handle things on their own"; on the other, SMEs are increasingly tight on money. Looking at these two things together, for business owners considering digitalization, it is a reminder that calls for sober reflection: the capabilities have arrived, but the accounts also need to be calculated clearly.
1. The Agent kernel has changed: from “following instructions” to “managing itself”
In the past, what we understood as an agent was more like an obedient executor: you give the steps, and it follows them. The Super-agent proposed by DeepSeek this time liberates the upper limit of multitasking from "human attention"—through event-driven and asynchronous multi-channel interaction, it allows multiple subtasks to advance in parallel, while an independent Watchdog monitors from the side. The team disclosed that an internal Super-agent based on this idea has already been running continuously, participating in problem reproduction, code review, and even its own iteration.
What it means for enterprises: When an Agent can self-manage and self-heal, its role in ERP upgrades from "helping you fill out forms" to "watching processes for you, spotting problems, and trying to solve them." This is exactly the technical prerequisite for the real implementation of "agentic ERP" that major ERP vendors have been talking about over the past year—the core must be stable enough before anyone dares let it touch business.
But the more autonomous the core is, the higher the demands it places on the company internally: you need to figure out earlier which processes can be allowed to "act on their own" and which must be decided by people.
II. During an economic downturn, digitalization must make the numbers add up
In October, the manufacturing PMI fell back to 49.3%, down 0.8 percentage points from the previous month. By size, large enterprises stood at 50.2%, medium-sized at 48.4%, and small at 47.5%—medium and small enterprises have been under continuous pressure. The new orders index was 49.4% and the production index 49.8%, both below the boom-or-bust line.
October manufacturing PMI, SMEs have already fallen below the boom-or-bust line
This dataset is a hard constraint for digital solution selection: During an economic downturn, what companies fear most is not "not being advanced enough," but "having deployed an asset-heavy system they cannot afford to use or control." The more this is the case, the more they should prioritize cleaning up the data foundation and start with an incremental, affordable approach, rather than trying to adopt the flashiest AI agents all at once.
For SMEs, the first goal of digitalization has never been "adopting AI," but rather "getting the books straight, understanding inventory clearly, and smoothing out processes." AI agents are amplifiers, not a lifeline.
III. The threat is not hypothetical: Agents have already been used to breach major companies
On November 1, the security team Cathedral publicly published a retrospective: they used an AI agent to breach Anthropic's internal systems within 12 hours, obtaining employee salary and internal communications information. Note that this was not a human hacker manually climbing over walls, but a group of autonomous agents "capable of scanning wall crevices tens of thousands of times per second," analyzing documents themselves, finding tiny vulnerabilities, and piecing together a breakthrough.
A blunt warning about ERP: An agent that can automatically read, write, and coordinate across multiple systems will likewise become a lever for attackers. ERP is the "general ledger" of enterprise data; once covered by agentic access, its exposure surface is far larger than that of traditional account systems. Least privilege, audit logging, and human fallback are not compliance decorations here—they are the bottom line.
This incident and the earlier industry survey finding that "58% of enterprises have deployed AI to access ERP data, and 22% have reported at least one instance of misuse" corroborate each other: when agents enter ERP, security risks arrive simultaneously—they are not something for the future.
IV. Pragmatic Options of Open-Source ERP for SMEs
In the trilemma of "wanting to use agents, yet fearing being co-opted, and also fearing being unable to afford them," open-source ERP offers a middle path. Take Odoo 19/20 as an example: its design orientation is that same-database integration lets context flow naturally, offline capability ensures business continuity, composability allows agents to be integrated as "pluggable capabilities," and 100% data ownership means you can take it with you at any time. With native MCP, enterprises can keep models, entry points, and governance in their own hands.
Controllable
Tiered permissions, auditable logs, replaceable models—governance rules are written in your own system, not in the vendor's black box.
Affordable
Start with the community edition at zero license fee, add agents and implementation on demand, avoid one-time heavy-asset investment, and align with cash flow constraints during a downturn.
Open source is not a "cheap alternative" but a sovereignty strategy: you use the same agent capabilities, but the keys, data, and context remain in your own hands. For SMEs, this is often more stable than chasing the most cutting-edge closed entry points.
V. Three Disciplines for Small and Medium-Sized Enterprise Owners
- Autonomy and governance advance in tandem.The more capable an agent is of working on its own, the more it must first establish permission tiers, log retention, and rollback mechanisms. First have brakes, then step on the gas.
- Least privilege + full audit trail. Each write permission granted to the Agent is scoped on a "just enough" basis; sensitive operations (payments, master data changes) require mandatory human confirmation, and all actions are logged.
- Use open source to reduce TCO in a controllable way. First use open-source ERP to clean up the data foundation and smooth out processes, then layer on agents as needed, avoiding being locked in by asset-heavy solutions and closed entry points.
Intelligentization is not a reckless gamble, but a gradual advance with brakes. The greater the capabilities, the clearer the boundaries must be; the more clearly the accounts are calculated, the more steadily the path is traveled.
Conclusion
DeepSeek teaches agents to learn to fix themselves, PMI reminds SMEs that money is tight, and the Anthropic incident shows that the threat has already become real. Stacked together, these three things give every enterprise preparing to adopt agents a simple conclusion: First clean up the foundation, draw the boundaries clearly, keep sovereignty in your own hands, and then let agents come in to work. No matter how fast technology runs, it should wait for your governance and ledger. Being half a step slow is often better than crashing.
This article is based on DeepSeek Super-agent technical disclosures, the National Bureau of Statistics' October PMI data, and a public retrospective on Anthropic being breached by an AI agent, and is written for SME owners and digitalization leads for reference; it does not constitute procurement or security advice.
Image: AI-generated (deep blue-cyan tech style) · Suitable for blog and Odoo official website publication
