Method · Causal Layered Analysis
Depth agents
Causal Layered Analysis, reimagined for the agent economy
Drilling past the headlines to the systems, worldviews and myths that actually drive change — repeatably, and across far more issues than a human analyst could hold.
The classical method
Causal Layered Analysis (CLA), developed by Sohail Inayatullah, reads an issue at four depths: the litany (the surface headlines), the systemic causes (economics, policy, technology), the worldview (the assumptions holding it in place) and the myth or metaphor (the deep story underneath). Strategy aimed only at the litany fails — it treats symptoms; lasting change has to reach the deeper layers, where the situation is actually held in place.
CLA's discipline is that it refuses to let analysis stop at the first, loudest layer. The latest scare story or product launch is the litany; underneath sit the structures that produced it, the worldview that makes those structures feel natural, and the metaphor — 'man versus machine', 'the rising tide', 'the arms race' — that quietly frames what everyone takes to be common sense. Intervene at the wrong layer and the system snaps back.
Why the human-run version hit a ceiling
CLA is demanding precisely because it asks one analyst to keep moving between very different levels of abstraction, holding the surface and the deep story in mind at once — cognitively expensive work that does not scale. In practice it gets applied to a few flagship issues in a facilitated workshop and almost nowhere else, and it is rarely revisited as the situation evolves, because re-running a four-layer reading by hand is as costly as doing it the first time.
How it works with agents
Depth agents do the layered reading systematically. For any issue, they separate the noisy surface from the structural drivers, surface the worldviews that make the situation seem inevitable, and name the underlying metaphor — then reason about how a change at each layer would ripple up to the others. The agent is well suited to this because moving fluidly between registers and reframing the same facts at different depths is, at bottom, a language operation.
Running it by machine makes it repeatable and comparable in a way the workshop version never was. The same four-layer reading can be applied across dozens of issues, compared side by side, and — crucially — revisited cheaply as the situation changes, so the deep reading stays current instead of fossilising in a workshop deck.
It also makes the method auditable. Because each layer's claims carry their evidence, a depth agent's reading is something a human can interrogate and contest layer by layer, rather than a charismatic facilitator's synthesis that is hard to reconstruct after the room has emptied.
What the evidence says
CLA is a structured reframing rather than a forecasting task, so the relevant evidence is about whether LLMs can hold and shift between frames — and the OECD–WEF AI in Strategic Foresight survey places exactly this kind of sense-making and reframing among the foresight tasks practitioners expect AI to augment. The deeper warrant comes from the demonstrated ability of these systems to sustain distinct perspectives and rewrite the same material across very different registers, which is precisely the cognitive move CLA formalises.
The honest framing is that depth agents are an instrument for structured interpretation, not for settling what is true. Their value is in forcing the harder questions and making the deep reading cheap and repeatable — not in delivering a verdict about which worldview is correct, which remains a human and political judgement.
Applied to the agent economy
Applied to the agent economy, depth agents stop us reacting to the litany — the latest model launch or scare story — and force the harder questions: which economic structures are actually shifting, which assumptions about work, agency and ownership are quietly breaking, and what deep story about humans and machines is being rewritten beneath the noise.
That matters because most of the public conversation about agentic AI lives entirely at the litany. A depth reading is how an institution avoids building strategy on a headline — and how it notices that the real action is a worldview shift (say, from software-as-tool to software-as-worker) that no single news cycle will ever name.
Where humans stay in command
The failure mode is a confident, fluent reading that imposes a tidy four-layer story on an issue that does not fit it — false depth. So every layer's claims must cite their evidence, and a depth agent's reading is treated as a provocation for human interpretation, not a finding. Where the agent names a worldview or a myth, that naming is explicitly contestable; a second agent and a human are invited to propose a rival reading.
And the choice of which deep story to act on stays human. Depth agents widen and sharpen the interpretation; deciding which layer to intervene at, and which metaphor an institution wants to help write, is a judgement an agent surfaces options for but never makes.
How we run it
- Litany — capture the surface narrative and headlines exactly as they circulate.
- System — identify the economic, technological and policy drivers underneath, each with its evidence.
- Worldview — surface the assumptions that make the situation feel fixed, and name the rival worldviews.
- Myth — name the deep metaphor, and reason about how a change at each layer would propagate upward.
- Contest — a second agent and a human propose alternative readings, and a human chooses where to intervene.