There is a 2025 technical development that, read carefully, says something old in new words. AI agent engineering —building software that no longer suggests but decides and executes— discovered that the factor separating a useful agent from a useless one is almost never the model. It is the context assembled for it: what it knows about your organization, in what vocabulary, with what memory of what already happened. The emerging craft of composing and maintaining that context has a name —context engineering— and it became the discipline that actually matters (Karpathy, 2025).
Read from my work, that is not an alien development: it is exactly what the bridge method already produces. Institutional memory —repeated questions getting consistent answers— and the project’s bilingual glossary —business and technical, defined once— are, literally, the context an agent needs to decide well. What is built in the first link, reading the organization before touching the technology, is the raw material any agent later thinks with.
The proposition with bite
Hence a sentence I use as a diagnostic test: the organization that cannot explain its process to a person cannot package it as context for an agent either. Organizational illegibility —the process that “lives in someone’s head”, the field each area understands differently, the real flow that does not match the formal one— used to be paid for in zero adoption: the tidy system no one used. Now it is also paid for in something more concrete: the impossibility of delegating. There is no context to give the agent, because it exists nowhere legible.
This inverts the usual order agents are sold in. The pitch is usually: buy the agent and then “give it context”, as if context were a file you upload at the end. It is the other way around. You do not start with the agent; you first build the legibility of the organization —what process does what, what each datum means, where the judgment lives— and that legibility is the context. The boring work of putting the house in order is not an annoying prerequisite: it is the asset itself.
Why this raises the stakes, not lowers them
It might seem that agentic AI makes the work of organizational reading obsolete —“let the agent sort itself out”—. The opposite happens. The more autonomous the system, the costlier bad context is, because the agent does not suggest an error someone can still discard: it executes a decision that has to be undone. And there is evidence that coordination does not automate itself: evaluated over twenty workflows, state-of-the-art manager agents still fail to jointly optimize goals, constraints and runtime (Masters et al., 2025). The sociotechnical reading —coordinating the human with the technical— is, for now, the residue that resists automation.
The bridge, then, does not only set the agent’s guardrails —the auditable rules, the escalation points, the revocation—. It does something earlier and deeper: it manufactures the material the agent thinks with. The legible organization is not a tidiness luxury; in the age of agents, it is the condition for being able to delegate anything to them without delegating in the dark.
The framework is developed in The sociotechnical bridge and in My approach. Related essays: The 95% that captures no value is not the models’ fault · Your team already adopted AI without you.