Invisible agents, not 10% efficiency, will win the assistant race
Key points
Key takeaways from the a16z show episode with Anish Acharya, General Partner at Andreessen Horowitz, and David Pawlan, creator of Assistant Benchmark (September 2026):
Four weeks of agent chaos. The personal-agent category exploded in the weeks before the recording: Poke (launched around 8 September, one of the first consumer AI agents, built around a price-negotiation gag), then OpenClaw in late November, and most recently Instinct, Grokbot and Muse, plus a long tail of tools like Caddy, Ali and Pi — 122 assistants now listed on Assistant Benchmark across categories.
How Assistant Benchmark works. Every assistant gets the same one-shot prompt (book a flight to Chicago, find a vegetarian restaurant within five blocks) and is scored on roughly 16 dimensions. Launched about two weeks before the episode, it had already drawn over 100,000 visitors and outreach from basically every founder in the space.
Cost savers beat time savers. The general public does not care about being 10% more efficient (a point tied to Ben Thompson's entertainment-over-productivity argument); the hook that lands is "free money" — agents that scan a year of receipts to file HSA reimbursements, or rebook flights to bank price drops as travel credits. One user's Grokbot wired to the weather system cut his sprinkler water bill by 50%.
Proactivity is the moat. The best assistant is invisible — checking you into flights, filing reimbursements, clearing your inbox before you sit down — with massive defensibility around proactivity. But autonomy sits on a fine trust line: cross it once, argues the conversation, and you lose the user permanently.
Surfaces and hardware fragment by user. iMessage versus standalone app is largely personal preference, with generational and gender splits predicted (Gen Z texting versus app interfaces; a visual dream-board for some, a functional thread for others). Hardware is the long shot: David's hot take is that Meta's Muse Charm is less a play to win hardware than a real-world data-collection device feeding the company's mapping and metaverse ambitions.
Travel and finance end up as connectors. Infrequent, high-value categories like travel are a strong viral hook but look likely to fold into horizontal agents as connectors rather than stand-alone winners. Finance carries the most administrative overhead for consumers yet is the least practical use case right now (agents won't reliably "make you a million").
Agents will talk to agents. Agent orchestration — agents finding the most efficient agents and delegating to each other, including development and design agents — is already one of the most-discussed use cases, and the two expect new infrastructure and products built purely for agent-to-agent interactions.
Read more: a16z on YouTube
