Chamath's Three Phases of AI: Why Context Is the Next Battleground

In this All-In Podcast clip, Chamath Palihapitiya argues that the second phase of AI, harnesses and agents, is ending, and the third phase belongs to whoever controls the contextual data that turns a generic agent into a trained professional. Models gave us raw intelligence. Agents gave that intelligence eyes, hands, and memory. The next step is teaching the agent to do a specific job, and that requires the proprietary context held by large systems of record.

The Three Phases

Chamath lays out a simple progression, using the brain as his running analogy.

Phase	What it is	His analogy
1. Models	Foundation models good at Q&A	A brain
2. Harnesses and agents	Scaffolding that makes models act autonomously	Giving the brain eyes, hands, a notebook for memory, and a keyboard
3. Context and specialization	Informed agents trained for specific jobs	Training that body to be a lawyer, customer service rep, or sales agent

His claim is that phase two is closing now. He says he already sees the shift in the enterprises he works with: an autonomous agent alone is insufficient.

Why Agents Alone Are Not Enough

An agent can browse, type, and remember, but it starts as what Chamath calls a primordial object that was good at Q&A. Making it useful for real work means giving it the background a professional carries: the account history, the case files, the deal pipeline. Without that, the agent does a generic job. With it, the agent does the job intelligently. The gap between the two is contextual information, and lots of it.

Systems of Record Hold the Leverage

The companies that store that context today are the large systems of record, platforms like CRMs and ERPs that hold customer, financial, and operational data. Chamath says these companies "hold an incredibly special place in the ecosystem if they do it right." He had made this call before and doubles down on it here. The conditional matters: the advantage only accrues if they open that context to agents rather than sitting on it.

The "Mark" Playbook

Chamath credits a CEO he calls Mark, in context the head of a large system-of-record company, with executing this transition well: playing nicely with the model providers, supporting third-party harnesses, building his own agents, and now pushing into the context phase. He ties that strategy directly to the financial results: strong net dollar retention, strong revenue, raised guidance, and a stock he says is up roughly 43% since he called the bottom. The strategy lesson is that incumbents with data do not need to win the model race. They need to make their data the mandatory ingredient for everyone else's agents.

Key Takeaways

• AI has moved through models (phase one) and harnesses plus agents (phase two); phase three is context and specialization.
• An autonomous agent without domain context is generic. Training it into a lawyer, sales rep, or support agent requires large amounts of proprietary contextual data.
• Systems of record control that data, which gives them outsized leverage in the agent era if they act on it.
• The winning incumbent playbook: cooperate with model providers, support external harnesses, build your own agents, and monetize your data as context.
• Chamath points to strong net dollar retention, revenue growth, and raised guidance as evidence the strategy shows up in financials.