Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board

Raw markdown memory outperforms polished forgetful apps

While mainstream consumer AI apps still present blank slates with no memory, experimental projects like OpenClaw achieve better continuity by writing messy notes to markdown files, creating a 'Memento'-style memory system that proves more effective than vector databases.

Code repositories provide ideal agent environments

Unlike general business tasks that scatter context across systems, code repositories concentrate textual context, test feedback, and formal change history in one place, creating an environment uniquely designed for robotic self-reflection and iteration.

File-based context enables true random access

Taylor argues that loading context from a directory of markdown files provides a more useful mix of structured and random access than vector databases, which require knowing what to search for upfront.

Engineers must abandon emotional attachment to code

Taylor describes forcing himself to stop caring about the elegance of handwritten code, recognizing that future software engineering prioritizes correctness through AI agents rather than craftsmanship of raw artifacts.

Documentation becomes the primary engineering output

As AI agents generate implementation code, the durable value shifts to documentation artifacts capturing intention and customer problems, potentially forcing engineers to focus on the very task they historically avoided.

Harness engineering defines the new workflow

The emerging discipline involves building context-rich environments—combining documentation, tests, and rules—around agents rather than writing code directly, though the field is still defining its fundamental categories and tools.

Agent harnesses will supersede traditional APIs

Beyond REST endpoints, companies will need to provide comprehensive 'harnesses'—instruction manuals and context layers—that teach agents how to extract maximum value from business objects.

Existing infrastructure outperforms new protocols

In healthcare, Sierra's AI agents already communicate via 'English over PSTN' (phone calls) rather than waiting for perfect API integrations, demonstrating that AI can effectively leverage legacy rails like the telephone network.

Multi-agent MCP architectures show limitations

Taylor argues that stuffing context into subagents creates robotic orchestration layers, suggesting that true agency requires monolithic context sharing more akin to OpenClaw's markdown approach than modular MCP server architectures.