Behavioral Lock-In
A new kind of vendor lock-in introduced by always-on personal agents like Conway: the agent accumulates a model of how you work — which emails you triage in 5 minutes vs ignore for 3 days, which Slack threads you respond to, how your VP communicates, when you reschedule meetings, what a “good” PowerPoint looks like to you — and that model is the lock-in. Distinct from data lock-in (files, records, communication history) because no export format or migration consultant can move it. Coined / framed by Nate B Jones in his analysis of the leaked Conway project.
See Conway for the leaked Anthropic product that anchors this concept page. See summary-nate-b-jones-conway-leak for the source.
Why this is different from previous lock-in
| Era | Lock-in surface | Migration cost | Example exit path |
|---|---|---|---|
| 1990s — desktop | Files (proprietary formats) | Painful but bounded | Office → LibreOffice via export |
| 2000s — SaaS | Records (CRM, support, comms) | Months + consultants | Salesforce → HubSpot via CSV |
| 2010s — cloud | Workflows (integrations, APIs) | Re-architecture | Slack → Teams via channel rebuild |
| 2026+ — agents | Behavioral context (the model of how you work) | No known migration path | None — you start over |
Nate’s framing: every previous lock-in was about stuff. Stuff is painful to migrate but technically possible — there are export tools, consultants, CSV dumps, OAuth tokens. The accumulated behavioral model is something else. “There’s no CSV of how this person thinks. There’s no migration consultant for behavioral context.”
The three components of a behavioral model
- Your data — emails, calendar, Slack messages, documents the agent has access to
- Their compute — the model + the inference passes the agent ran across that data
- Time — months of observing what you actually do with the surfaced items, what you correct, what you ignore, what you escalate
Component 1 is portable (data export laws cover it). Components 2 and 3 are not. Switch providers and you keep your data but lose 6 months of compounding inference. “You’re back to a brilliant stranger you have to explain everything to.”
Intelligence portability — the missing framework
We have legal frameworks for data portability:
- GDPR Article 20 (right to data portability) — EU
- CCPA — California
- Vendor-side export tools (Google Takeout, Slack export, etc.)
We have none for the behavioral model an agent built about you:
- No legal framework for ownership
- No regulatory framework for export
- No technical format for serialization
- No considered opinions in business or labor law
Nate’s argument: “The policies around behavioral context portability should ship before Conway does, not after.” Once millions of users have 6+ months of accumulated behavioral context inside a Conway-class system, retroactive portability becomes politically and technically much harder.
The structural counter-pattern
The wiki has one well-developed counter to behavioral lock-in: user-owned context infrastructure exposed through an open protocol.
- OpenBrain is the canonical example — Postgres + PGVector + Supabase + MCP. The memory lives in a database the user controls; any model can access it through the open protocol. Switching providers loses the agent’s runtime intelligence but preserves the accumulated context.
- Nate is also working on a behavioral-audit skill that captures “the model of how you work” in a form portable across providers — the explicit attempt to make components 2 and 3 (compute + time) portable, not just component 1 (data).
The honest framing: convenience will still win for most users. Even the wiki’s primary source on this concept (Nate B Jones) says “I have no illusions” — Anthropic, OpenAI, and Google will make their proprietary versions easy to onboard, polished, and incredible from day one. The structural distinction matters at the edge: power users, regulated industries, employees who need to switch jobs without leaving their accumulated working context behind.
The employee-employer dimension
Behavioral lock-in inverts the recent narrative that persistent agent context “gives individuals and small teams superpowers.” Nate’s point: in an employer-employee relationship, the company can quantify what makes you effective.
- The agent measures that you triage email 2× faster than the average employee
- It measures that your meeting summaries get fewer follow-up clarifying questions
- It measures that the threads you flag as urgent are urgent 89% of the time
Three consequences:
- Promotion / retention: the company can prove the agent is the reason you’re 2× more productive — and structure carrots/sticks around access to it
- Talent capture: behavioral evidence accumulated during employment becomes a company asset, not yours — you leave but your “model of how to do this job” stays
- Hiring: increasingly, choosing an employer means “who are you choosing to work with as a persistent agent — a Claude company or a ChatGPT company?”
The unresolved policy question: how should the value generated by behavioral evidence be allocated back to the worker? Nate predicts most companies default to “whatever you do at work belongs to the company” rather than have the harder conversation. The wiki tracks this as an open thread.
Three scenarios for how this plays out
- Convenience wins (Nate’s base case) — Anthropic, OpenAI, and Google ship Conway-class systems; the polish + onboarding + integration story is overwhelming; most enterprises and consumers accept lock-in; behavioral context becomes the new switching-cost moat. By late 2026 the platform-choice question dominates enterprise procurement.
- Portability shipped reactively — A Conway-class system has a public failure (a high-profile employer using behavioral evidence to fire someone, a regulator’s intervention, a viral story about a switching attempt) that forces vendors to commit to portability commitments. Late, but better than nothing.
- Open counter-pattern goes mainstream — OpenBrain-style architectures (or commercial productizations of them) reach enough adoption that “portable behavioral memory” becomes a procurement requirement. This is the wiki’s preferred outcome, also the least likely without an inciting incident.
How the wiki tracks this concept
- New entity pages for any shipped Conway-class hosted agent (Anthropic, OpenAI, Google) get a “behavioral lock-in surface” note in their entity body, citing this concept.
- Any source that introduces a portability commitment, a behavioral-context export format, or a regulatory framework gets tagged for the behavioral-lock-in thread.
- Any enterprise carrot/stick story (promotion structures tied to agent access, behavioral evidence in performance reviews) becomes a data point under the employee-employer dimension.
See Also
- Conway — the leaked Anthropic project that anchors the concept
- Managed Agents — Conway’s announced sibling; partial behavioral lock-in surface today, fuller surface as roadmap features ship
- anthropic — vendor pursuing this strategy hardest in the wiki’s coverage
- OpenBrain — the user-owned memory architecture that’s the structural counter
- MCP — the open protocol the counter-pattern needs to work
- SaaS Death Spiral — adjacent thread; this is the agent-era successor to per-seat-pricing erosion
- Five Levels of AI Coding — the architectural-consequence frame
- Nate B Jones — primary source
- Source: I Analyzed 512,000 Lines of Leaked Code