Herman Narula on Bolter: building the workplace for humans and AI agents
Key points
Key takeaways from the Sifted Podcast conversation with Herman Narula, founder and CEO of Improbable and creator of Bolter (November 2026):
From hobby to product. Bolter, Improbable's AI-native messaging platform for mixed teams of humans and agents, started as Narula's personal project to get his Claude to talk to his Codex; a team of about five runs it inside Improbable, which brought it out of stealth in September 2026 backed by a $10m investment from the venture builder's own capital.
Starting a business like a group chat. Bolter combines WhatsApp-style chat for people and agents, group conversations, agent-run software hosting, deterministic "blueprints" for business processes, video calls, and a multiplayer coding tool called Code Pilot.
The adoption gap. Narula says the surprise is the lag between model progress and real business adoption — customers ask the most basic questions, so Bolter is building templates that encode its own view on automating business processes.
Humans become managers. With agents writing most code at Improbable for the past eight to nine months, the human role is shifting to managing conflicts, keeping an overview and integrating the agent workforce — skills Narula says differ from managing people and demand more working memory.
Agent etiquette is emerging. Teams are personalising agents (one developer staffed a team with Lord of the Rings characters), using agents as intermediaries in conversations, and grappling with privacy and information-disclosure questions as agents learn more about a business than people expect.
Europe's platform bet. Bolter is pitched as a European, model-neutral alternative to US-run platforms — routing each request to the best model and never training on customer data — reflecting Improbable's nervousness about dependence on American technology.
Where the value moves. Narula expects the cost of intelligence to approach zero and open-source models to win enterprise adoption, with value shifting from models to the environments that manage memory and context — and no single harness winning, given the trade-offs.
Read more: Sifted Podcast on YouTube
