Will software need to be built, sold, and used for agents rather than human users?

Jerry Murdock · Feb 28, 2026

Autonomous agents already write software faster and cheaper than any human being, so companies that don't build their software for autonomous agents will be severely challenged

If your software is made for agents and they have a reason to use it, you'll do fine; if you still believe human beings will buy your software, you're in trouble

Scope: timing uncertain — 'maybe six months, maybe a year'; expects most enterprises to take a year or longer to commit to autonomous agents

24:37 20VC: Why Cursor is Dead | An AI Tsunami is Coming & You Need to Prepare | Systems of Record Become Valueless Databases with Agents | Is This The End of Tech Private Equity with Jerry Murdock, Co-Founder of Insight Partners

Aaron Levie · Apr 20, 2026

The mandate for software companies now is to own the workflow and the data and be the place agents intentionally choose to work with that data; anything else is a tough spot

Agents need a best destination for category-specific data and workflow automation, so platforms that aren't that destination lose relevance

40:41 20VC: Everyone is Wrong; We Will Have More Developers in Five Years | Why Frontier Labs Will Be Way More Valuable Than They Are Today | Are SaaS Companies Cooked: Which Thrive & Which Die with Aaron Levie, Founder at Box

Jerry Murdock · Feb 28, 2026

Software will shift from being purchased by human beings to being purchased and used by agents, with autonomous agents functioning as employees who are given credentials and identity and reviewed on what they bought, spent and accomplished

The agent will make the purchasing decision itself, so you manage it like an employee

22:08 20VC: Why Cursor is Dead | An AI Tsunami is Coming & You Need to Prepare | Systems of Record Become Valueless Databases with Agents | Is This The End of Tech Private Equity with Jerry Murdock, Co-Founder of Insight Partners

Julien Bek · Aug 24, 2026

Agents are the new customer: on the demand side there is a new class of customer that businesses currently serve far worse than human customers

Three years into AI we're already at parity between agent and human traffic; today agents delegate out of convenience but as models get smarter delegation becomes rational because a 500-IQ AI simply makes better decisions

Scope: framed as 'my thesis'

47:18 20VC: Inside Sequoia's Investment Committee: Lessons from Don Valentine, Doug Leone and Alfred Lin | How the SpaceX and Citadel Deals Went Down | What Sequoia Specifically Looks for in Founders with Julien Bek

Aaron Levie · Apr 20, 2026

Unstructured data is where agent tool usage will most exceed human usage, and we are about to see an explosion of unstructured data.

Agents will read every contract and generate contracts, marketing assets and reports — making creation and review trivially easy — while you still need a backbone to manage, coordinate and set guardrails on those workflows.

17:57 Unstructured content explosion needs a governance backbone

Aaron Levie · Apr 20, 2026

Being the best place for agents requires not just clean APIs but pricing models and surrounding compliance capabilities around the API

For example an agent working with FINRA-compliant documents needs clean APIs plus retention and audit features so the customer can prove compliance to regulators — that combination is why they'd build on Box

42:02 Compliance and pricing around the api not just clean apis

Aaron Levie · Apr 20, 2026

Software is going headless and platforms must be headless-first — a view he has become fully convinced of in the past year

Agentic tool-calling and cross-system search accuracy improved far faster than expected; two to three years ago an agent told to find and process a document in Box would almost always pick the wrong file and couldn't read it, and that has now changed

42:54 Headless first platform design is now mandatory

Nikesh Arora · Jun 22, 2026

The only logically viable way to do agentic security is to aggregate agent traffic through a gateway, firewall or router so you can observe and stop agents

If enterprises agentify, you need to know how many agents are running, track them, govern them and secure against them, which requires a single chokepoint for traffic

48:06 Agent traffic must route through a gateway to be governed

Jerry Murdock · Feb 28, 2026

Infrastructure built for agents has qualitatively different performance requirements than infrastructure built for humans — latency below human perception thresholds still matters because agents notice it

Humans only perceive delays around 400 milliseconds, which is where most systems and sandboxes are tuned, but E2B's sandboxes respond in milliseconds and when an agent spins up sandboxes in seconds that response time is critical

17:03 Agent grade infrastructure needs differ from human grade

Elias Torres · Mar 21, 2025

Software of the last 20-30 years was built assuming a human operator, and the future is software built with no operator at all — driverless-car software rather than cars needing drivers

Users don't want to learn to drive the tool; they want to say where and when they want to go and have the LLM do the work

16:30 Software shifts from human operated tools to autonomous driverless systems

Matan Grinberg · Jun 13, 2026

Within about five years, we will look back in disbelief that highly paid engineers spent hours writing release notes and documentation by hand

Human DevX impact scales linearly with headcount, but agent DevX impact scales with number of agents; better DevX means agents adhere to standards, reducing review burden and increasing throughput

35:01 Developer experience must be built for agents because it scales with agent count

Matan Grinberg · Jun 13, 2026

The value of a great API is increasing as agents become buyers, because what makes systems easy for agents is currently the same as what makes them easy for humans

Agent-friendliness and human-friendliness currently coincide

38:02 Agent needs currently coincide with human needs so good apis suffice

Julien Bek · Aug 24, 2026

Agents are not just a new category but a whole parallel economy, within which new categories like AEO will be created

50:37 Agents form a parallel economy spawning new categories

Harry Stebbings · Aug 24, 2026

Large enterprises will not shift to agent-driven buying as fast as the market expects, because enterprise purchasing cycles are slow

People overestimate what happens in a year and underestimate what happens in ten; enterprises don't buy quickly and mostly don't even use tools like Slack

52:36 Enterprise purchasing inertia delays agent buying

Matt Swulinski · Aug 15, 2026 · hedged

Companies that are best at PLG today will also be best positioned for the agentic layer, because optimizing a self-serve funnel for humans and optimizing for how agents research and select tools requires the same framing and diligence.

PLG assumes a perfect self-serve funnel to the human; the same discipline applies to understanding how agents make tool and API selection decisions without a human in the loop.

5:04 Plg funnel discipline transfers to agent selection