We're in the "agents as employees" phase. It won't last.

Most teams are still defining agents as human roles. The recruiting agent. The customer support agent. The marketing agent.

That framing makes sense right now. It maps to how organisations already think about work.

But it’s a transitional model, not the end state.

Here’s what I think is actually happening:

Work is being decomposed into tasks and workflows, not job titles.

A human does multiple things. An agent shouldn’t mirror that - it should execute a well-scoped unit of work within a designed loop: scope, execute, review, evaluate, deploy.

Garry Tan open-sourced his Claude Code workflow recently. The insight isn’t the tooling. It’s that the performance gain comes from how the work is structured - task boundaries, review gates, iteration discipline. Not prompt magic.

The bottleneck has shifted from building to reviewing.

Throughput is no longer the constraint. Judgment is.

The best agent-native teams I’ve seen all converge on the same design choice: human-in-the-loop approval on outputs, learning loops from human edits, and evals in CI to catch regressions.

They don’t just automate work. They redesign the quality control surface.

When you increase production speed without increasing review capacity, you don’t ship faster. You ship worse.

Context, memory, and self-learning are becoming infrastructure.

Agents fail not because models are weak, but because they operate on stale context, have no persistent memory, and can’t learn from their own mistakes. They hallucinate APIs. They forget what they learned last session. They repeat the same errors.

The fix isn’t a better model. It’s curated docs, persistent memory, eval frameworks that catch regressions, and feedback loops that make agents smarter over time. Andrew Ng’s Context Hub work points at exactly this.

Context, memory, and self-learning quality now determine output quality more than model selection does.

Today we model agents as employees because that’s what companies understand.

Tomorrow, work gets abstracted into programmable workflows where the interaction layer is structured contracts between systems and agents. Browser automation is still brittle. Middleware adapters are band-aids.

And the handshake layer between agents and systems? Still unsettled. MCP is already being challenged. Others argue skills-based patterns will win over protocol-based ones. The truth is we’re in the messy middle of figuring out how agents talk to systems and each other.

That’s fine. The protocols will shake out. What won’t change is the underlying shift: agents need native, well-defined interfaces - not screen-scraping and prompt hacks.

The teams that win won’t have the most agents. They’ll have the best workflow architecture, the strongest review discipline, and context.