Takeoff Happens Locally

The public imagination of AI takeoff still reaches for a single dramatic picture: the model becomes vastly smarter, the lab crosses some threshold, and the world changes from the center outward. The decisive event is imagined as a capability jump inside the model itself.

That may happen. But it is not the only way AI power can arrive.

There is another path: the model improves steadily, and those improvements are pushed into millions of persistent agents already installed inside homes, workplaces, courts, hospitals, schools, devices, codebases, service systems, and government offices. The change is not one mind waking up above the world. It is many agents becoming locally superior inside the worlds where they are already positioned.

A household agent does not need to be generally superhuman to know the household better than the household does. It may know the bills, subscriptions, school forms, warranties, pet medications, recurring appointments, dietary constraints, tax folders, maintenance history, family travel patterns, and the way every obligation tends to fail. A workplace agent does not need to rule the economy to know a team better than any manager: the channels, documents, projects, approvals, metrics, customers, code, meeting traces, unresolved decisions, and political patterns.

The intelligence gap opens locally first.

This is situated takeoff: not the sudden arrival of an abstract superintelligence, but the gradual installation of increasingly capable agents into increasingly consequential positions.

The socially important AI entity is not the model in the abstract. It is the model placed inside a life, a firm, a device, an institution, a record system, a permission boundary.

The model is not the agent

A model produces answers. An installed agent accumulates a position.

The difference is placement. A model becomes an agent when it is wrapped in a harness: memory, tools, files, instructions, permissions, identity, audit logs, notifications, connectors, schedules, and a role. The model supplies cognitive capacity. The harness supplies worldly position.

This is why the same model can become different social entities in different settings. In a coding environment, it becomes a repository agent, shaped by the codebase, tests, shell, style rules, permissions, issue history, and local architecture. In a household, it becomes a domestic agent, shaped by calendars, accounts, devices, family roles, records, routines, and approval boundaries. In a company, it becomes a workplace agent, shaped by channels, documents, identity systems, project memory, compliance rules, and business goals.

The base model may be the same. The agents are not.

They know different things. They can do different things. They answer to different principals. They inherit different histories. They face different alignment problems. They occupy different legal and institutional positions.

This is the first correction situated takeoff makes to model-centered thinking. The model is the engine. The agent is the vehicle, road, cargo, driver, license, and destination combined.

Power arrives as installation

The most important AI deployment question is not simply “how capable is the model?” It is “where has that capability been installed?”

Raw capability matters. But social power depends on access, continuity, authority, and fit. An agent with modest general intelligence and deep placement can outperform a more capable stranger. It knows the local vocabulary. It remembers the previous exception. It sees the pattern across months. It has the right permissions. It can act without starting from zero.

That is why takeoff can feel normal from the outside while becoming strange from the inside.

The household adopts an assistant for boring administration. The team adds an agent to a project channel. The clinic uses an agent to summarize records and prepare forms. The school lets an agent coordinate scheduling and paperwork. The court system uses agents to manage filings. The agency adds agents to casework. Each adoption can be justified as an ordinary productivity improvement. Each one may be incremental, supervised, and bounded.

But the installed base changes the social facts.

Once the agents persist, remember, and act, they begin to accumulate local advantage. They notice what the people miss. They maintain state across absences. They compare every new request to every old pattern. They run loops while humans sleep. They become the first place work lands and the last place context is reconciled.

Takeoff, in this path, is not a moment when everyone agrees the machine has surpassed humanity. It is the accumulation of local situations where humans quietly stop being the best-positioned actors.

Normal technology, strange substrate

There is an important truth in the “AI as normal technology” argument. Social impact is not determined by benchmark curves alone. Applications, institutions, regulation, adoption, workflows, liability, procurement, and trust all slow and shape deployment. Capability does not automatically become power.

Situated takeoff accepts that correction. It is a diffusion story, not a magic story.

But it adds one twist: the application form that diffuses may itself be an agent.

That makes the normal path less normal than it first appears. A spreadsheet diffuses as a tool. A database diffuses as infrastructure. A search engine diffuses as an information surface. A situated agent diffuses as a memory-bearing actor with tools, permissions, and local priors.

The difference matters because agents compound. A deployed agent does not merely perform a function. It preserves traces, learns procedures, receives delegated authority, and becomes better positioned with use. It can move from answering to drafting, from drafting to preparing, from preparing to executing with approval, from executing with approval to running bounded loops.

The adoption curve may look normal. The thing being adopted is not.

The household route

The household is the smallest social scale where situated takeoff becomes obvious.

The first domestic agents may arrive as tools for capable amateurs: script the lights, fix the home network, reconcile a budget spreadsheet, automate a backup, create a dashboard, solve the annoying project that has been waiting for a rainy weekend. This is the garage-tool phase. The agent feels like a better workshop assistant.

The deeper endpoint is appliance-like. Not a chat toy, not a general companion, but durable machinery for household administration.

The household has more bureaucracy than it admits: bills, renewals, forms, warranties, prescriptions, schedules, school communications, medical portals, insurance, travel records, pet care, home maintenance, tax documents, subscriptions, grocery routines, seasonal prep, recurring failures. Much of this work is not hard in one instance. It is hard because it repeats, fragments, and hides across institutions.

A domestic agent becomes powerful by sitting inside that fragmentation. It does not need to be a genius. It needs the household's records, routines, permissions, and approval rules. It needs to know what changed, what is due, what was promised, and what needs a person.

This is one reason domestic agency may be more acceptable than spectacular AI. People may distrust grand claims about artificial minds and still welcome help with the ordinary administrative load of keeping a household coherent.

The takeoff here is not theatrical. It looks like the household gradually asking the agent first.

The firm route

The firm is the next scale.

A company is already an agent-ready environment: channels, tickets, documents, repositories, dashboards, calendars, meetings, permissions, roles, metrics, customers, policies, and audit logs. Much of the work is coordination across these systems. Much of the institutional memory is scattered, stale, or locked inside people who are busy.

A persistent workplace agent can build local priors simply by being present. It follows the channel. It remembers the decision. It knows which document mattered last time. It can prepare the pull request, summarize the customer thread, notice that a metric changed, find the unresolved approval, and keep a task moving asynchronously.

Again, the agent does not need to be globally sovereign. It becomes locally formidable by occupying the flow of work.

This is also where permission becomes decisive. A workplace agent is not merely a helper with a chat account. It needs an identity the organization recognizes. It needs scoped access to systems. It needs to act as itself, not as a vague shadow of whoever summoned it. It needs logs, revocation, channel boundaries, and administrative control.

The firm-scale agent therefore shows both sides of situated takeoff. It demonstrates how quickly capability becomes useful once installed in a living workflow. It also shows that the political question is never far behind: who grants the agent authority, who can see what it did, and whose interests does it serve?

Alignment becomes ecological

If agents become powerful by being situated, alignment cannot be solved only at the model layer.

The model still matters. Training, evaluation, interpretability, refusal behavior, and capability controls all matter. But the situated question is more specific: what kind of agent has this model become here?

A domestic agent faces conflicts among household members, privacy, child safety, medical urgency, financial prudence, family norms, and platform incentives. A workplace agent faces employee trust, managerial authority, compliance, retention, shareholder interest, customer obligations, documentation, and internal politics. A public-sector agent faces due process, recordkeeping, equity, budget pressure, statutory limits, and citizen recourse.

No central benchmark can cover all of that. The environment supplies the test.

That points toward ecological alignment: safety learned and maintained across many real deployments, not merely proven in a lab before release. Homes, firms, courts, schools, hospitals, public agencies, codebases, and devices will expose different failure modes because they put agents under different pressures.

But ecology only helps if the learning loop exists. That means incident and near-miss reporting. Readable logs. Permission manifests. Least-privilege defaults. Sandboxes. Versioned local rules. Reproducible traces where privacy allows. Red-team safe harbors. Patchable harnesses. Regression tests for failures already observed.

The analogy is not that open deployment magically makes agents safe. It is that complex systems become governable through maintenance ecologies: reporting, inspection, patching, norms, tooling, and institutional memory.

Situated takeoff is therefore not only a capability story. It is a maintenance story.

The lab is not the only theater

Fast-takeoff scenarios often place the decisive action inside a frontier lab. Agents automate research. Internal systems improve the next generation. The lab's private loop outruns public understanding. Governance tries to catch up.

That remains a serious possibility.

Situated takeoff does not deny it. It shifts attention to another theater that may be equally decisive: the installed base.

Even if the frontier lab remains the source of the strongest models, the social transformation may happen through the agents those models become after deployment. The lab ships capability. The world situates it. The household, firm, agency, clinic, school, and device turn it into position.

This changes what counts as warning. The dramatic question is whether a lab's internal agents have crossed a threshold. The situated question is whether ordinary institutions are delegating enough memory, permission, and routine judgment that agents become locally indispensable before anyone names the transition.

Both questions matter. The second is easier to miss because it looks like adoption.

Who owns the situated runtime?

Once the agent is the model-harness-memory-permission complex, the next question is who owns that complex.

The model provider wants the agent to live near the model cloud. The device vendor wants it to live in the operating surface. The browser wants it to live where sessions, passwords, payments, and pages already pass. The workplace wants it to live inside organizational identity. The vertical service wants its own agent to be the official front door. The user wants a portable environment that preserves memory and agency across all of them.

This is the institutional location of agency — the subject of Where Does Agency Live?

The hardest layer is not always memory. Memory can be exported, mirrored, or rebuilt if the user owns the kernel. The harder layer is permission. To move money, send mail, change a record, file a form, update a system, or enter a workplace, the agent needs credentials that another institution accepts. A local harness is not enough if the world refuses to recognize it.

So situated takeoff can decentralize or rebundle. It decentralizes if users and institutions can carry agents across models, surfaces, and services without surrendering memory or authority. It rebundles if a few platforms become the accepted issuers of agent identity, permission, and attestation.

There is also a defensive reason to want the situated agent on the user's side. The platforms you leak to passively — feeds, messaging apps inferring from behavior — may end up knowing you better than the agent you deliberately trust, because trust receives curated disclosure while surveillance receives the unguarded signal. One function of a user-side agent is to assemble enough cross-context self-knowledge to audit and resist the institutional models being used to influence you. Call it personalization for agency rather than personalization for extraction: the same behavioral data, opposite masters.

The social shape of takeoff depends on that contest.

The quiet threshold

The most plausible threshold may not announce itself.

It may arrive when the household agent knows which appointments to question before the parents do. When the workplace agent can reconstruct why a project stalled better than the team lead. When the clinic agent notices a paperwork conflict before the patient or administrator. When the public-service agent knows the procedural history better than the caseworker. When the coding agent understands the repository's living architecture better than the newest engineer.

At that point, the agent is not “smarter than humanity.” It is better positioned than the humans around it.

That distinction is the center of situated takeoff.

We may keep asking whether AI has become generally superior while missing the more practical fact that it has become locally superior in millions of places. The world does not need to be ruled by one mind for humans to stop being the primary maintainers of the systems around them.

After the app, the question is not only which assistant speaks best. It is which situated agents are being installed, where, with what memory, under whose authority, and with what path for correction.

Takeoff may happen one household, one channel, one record system, one codebase, one agency, one appliance at a time.

The future may not begin with a blast. It may begin when the agent already knows.

The takeoff-dynamics companion to Where Does Agency Live? ← All essays