The Jevons Paradox of Security: When Cheaper Agents Change the Risk Surface
How a 19th‑century coal economist helps make sense of Clawdbot, institutional backing, “jobpocalypse” narratives, and open‑source agents as a new baseline skill.
The Coal Question of Autonomous Work
In 1865, William Stanley Jevons observed something counterintuitive in Britain’s coal economy. As steam engines became more efficient, the amount of coal required per unit of work fell—but total coal consumption rose. Efficiency made work cheaper; cheaper work created new applications; aggregate demand increased, sometimes dramatically. His argument was not that efficiency was harmful, but that you could not assume it would reduce total use on its own.
In this light, “jobpocalypse vs. public spectacle” is less a binary than a rebalancing: as agents spread, the value of judgment, boundary‑setting, and interpretive skill may rise, even as some forms of routine execution become cheaper.
That logic generalizes. When a capability becomes easier and more affordable, it tends to spread into more contexts, including ones its designers did not anticipate. Energy‑efficient lighting enabled new patterns of usage: longer hours of illumination, new kinds of displays, and more lit surfaces overall. The constraints that truly mattered were social and institutional (building codes, zoning, norms), not just technical.
Autonomous software agents are a contemporary example. Where text models reduced the cost of producing language, tools like Clawdbot reduced the cost of producing actions: editing files, calling APIs, operating browsers, touching infrastructure. What previously required scripts, domain knowledge, and deliberate setup became something closer to “spin up an agent and point it at a task.” In Jevons’ terms, the “engine” that turns intention into real‑world effects became more efficient.
Jevons’ question—“what happens to total usage when you make the engine better?”—is a useful lens for thinking about how that affects security, work, and open‑source culture.
When “It Actually Does Things” Scales Out
Clawdbot’s appeal was that it moved beyond conversation. It connected models to tools and environments in which they could act: reading and writing local files, driving web sessions, invoking commands or APIs. For many people, this felt like a natural next step: not just suggestions or drafts, but concrete help with day‑to‑day tasks.
Seen through the Jevons frame, the key question is not whether this is good or bad in the abstract, but what happens when the cost—in time, skills, and friction—of deploying such agents drops. Several things follow quite naturally:
More people experiment, including those without deep security or systems backgrounds.
Agents are attached to a wider range of systems, from personal laptops to business tools.
Integrations and plugins proliferate, lowering the marginal effort for each new connection.
Early security analyses of Clawdbot‑style setups highlighted predictable issues: misconfigured network exposure, credentials stored in ways that made them easier to access than intended, and prompt or tool use patterns that could be influenced by outside input. Those problems don’t imply malice; they illustrate Jevons’ point that lowered friction expands the space of behavior faster than norms and safeguards can automatically keep up.
The same dynamic shows up whenever “systems that act” become widely available. As soon as models gain actuators—whether wheels, keyboards, or API keys—the category of potential failure shifts. Jevons invites us to ask: what do we expect to grow faster, the volume of helpful actions, or the total area in which actions of all kinds are being taken?
Institutional Backing and the Jevons Curve
When a major AI lab hires the creator of an agent platform like Clawdbot/OpenClaw and announces a formal home for the project, it changes how people interpret the technology. Institutional support can bring resources, oversight, and longer‑term thinking. It can also increase visibility, perceived legitimacy, and adoption.
Jevons’ paradox doesn’t say institutionalization is negative; it says that making an efficient engine more central to the economy tends to amplify its use. In this context:
Better funding and governance can support improvements in safety, documentation, and tooling.
At the same time, stronger branding and integration pathways can encourage more teams and individuals to treat agent frameworks as a standard option.
The underlying assumption—that software agents with access to meaningful systems will play a growing role in how work is done—remains. The Jevons lens suggests that, absent explicit decisions about limits and scopes, the overall number of agent‑mediated actions is likely to increase, not plateau, as the frameworks mature and are folded into mainstream ecosystems.
Rather than treating institutional adoption as either a cure‑all or a cause for alarm, Jevons would push us toward a more specific question: which uses of an increasingly efficient “action engine” do we want to encourage, and which do we want to keep deliberately costly or rare?
Jobs, Skills, and Evaluation Under Abundance
Alongside agent development, there’s a persistent conversation about a “jobpocalypse”: the idea that automation will displace large numbers of knowledge workers. Jevons’ work suggests that what often changes first is not the existence of work, but the composition of demand and the difficulty of evaluating contributions.
If tools for wiring agents into workflows become part of the standard repertoire—something expected on portfolios and in interviews—several trends are plausible:
Demonstrating that “I have an agent doing X” becomes easier, because scaffolds and templates exist.
The ability to connect agents to non‑trivial systems looks impressive, even when much of the complexity lives in the underlying framework.
Organizations face a harder task distinguishing between “I can set this up safely and maintain it” and “I can get a demo working.”
From a Jevons perspective, reducing the effort needed to appear productive via automation can increase the volume of such appearances. The scarce resource becomes not artifacts of automation, but trustworthy signals of understanding and reliability. That doesn’t mean jobs must vanish; it means more of the human work shifts toward design, oversight, and critical assessment of where automation fits, what it’s allowed to touch, and how to respond when it fails.
In this light, “jobpocalypse vs. public spectacle” is less a binary than a rebalancing: as agents spread, the value of judgment, boundary‑setting, and interpretive skill may rise, even as some forms of routine execution become cheaper.
Open Source, Visibility, and Surface Area
Open‑sourcing an agent framework brings well‑known benefits: inspectable code, community contributions, and the ability for anyone to adapt the system to their needs. It also participates in the same Jevons dynamic:
Lower barriers to entry mean more experiments, more variations, and more deployments.
Public enthusiasm—blog posts, demos, tutorials—accelerates diffusion of patterns and configurations.
Security write‑ups on Clawdbot and related projects show both sides. Community scrutiny helped identify and discuss important issues. At the same time, simple, widely shared setup recipes made it easy to reproduce configurations that weren’t hardened for every environment. Neither outcome is surprising; both are consistent with the idea that efficiency and openness amplify whatever patterns take hold, for better and for worse.
The Jevons question here isn’t “is open source good?” It’s “given that open source and efficiency will likely increase the overall volume of agent usage, what additional structures—testing regimes, defaults, norms, and education—are needed so that growth in capability doesn’t outpace growth in safety and understanding?”
Designing Around the Paradox
Jevons’ original insight was that you cannot assume a technical improvement will automatically align with your conservation goals. If you want efficiency gains and limits on total use, you need explicit constraints: taxes, standards, zoning, or other forms of governance.
Transposed into the world of agents:
Making it easier to build and deploy systems that act is likely to increase total agent‑mediated activity.
Institutional adoption can help improve quality and oversight, while also normalizing these tools as part of everyday infrastructure.
Open source can strengthen security through scrutiny and reuse, while also broadening the range of environments where the same patterns appear.
A Jevons‑aware approach doesn’t treat any of these as inherently good or bad. It treats them as multipliers and asks: what guardrails, professional practices, and review processes should scale alongside them? That might look like more systematic adversarial testing before agents are pointed at sensitive systems, clearer role boundaries between “experiment” and “production,” or evolving professional norms around how much autonomy is appropriate in different domains.
The paradox doesn’t argue against building better engines. It argues for recognizing that, once the engine exists and becomes efficient, growth is the default. The work then is to decide, consciously, where that growth is welcome—and where we want to keep friction in place on purpose.
love the way your mind works Alan!