Agents Expand Software, AI Engineers Build the App Layer
Agents make vast uneconomic software viable, surging engineer demand. Focus on practical archetypes like 24/7 ops and compressed research. Application layer on commoditizing models captures value—Europe leads here.
Agents Dramatically Expand the Software Universe
Agents fill the gap between software that should exist and what traditional coding could economically build. Previously, complex if-statements, business knowledge, and custom logic made much automation too expensive—think niche business processes requiring hardcoded rules. Agents change this: the full circle of desirable software becomes viable with minimal effort, leading to exponentially more software overall.
This drives the 'SaaS copocalypse' where companies shift from buying SaaS to building custom agents, even if faster. Result: higher demand for AI engineers, as cheaper software production tests market elasticity without hitting an S-curve yet. Engineers thrive because juniors excel in AI-native workflows, and veterans adapt—more software means more work, even at higher speeds.
On Vercel.com, AI agents already dominate: over 60% of last week's pageviews were from agents, shifting usage from UIs to APIs/CLIs. Demand CLIs for every feature; UIs are now cheap commodities.
Practical Agent Archetypes Deliver Immediate ROI
Skip hype-heavy coding agents; target low-hanging fruit for millions in savings without process overhauls. Four archetypes working today:
- 24/7 Operations: Automate jobs limited by human sleep. E.g., agent-as-a-service for support (90% deflection rate at Vercel's in-house agent), handling routine queries like credit card issues, boosting team satisfaction by eliminating toil.
- Compressed Research: For event → research → human decision flows, agents handle research only. Cuts 30min to 5min per instance; at 100k/year scale, massive savings. Vercel examples: contact sales agent (75% routes to support after LinkedIn/Google checks; saves 15min/person); abuse reports agent pre-researches sites.
- Surface Existing Info: Pull scattered data (Slacks, Granola recordings, issue trackers) into usable form. E.g., auto-update issue trackers or pre-compile manager update lists—info exists but was practically inaccessible.
- Eliminate Hated Tasks: Ask teams what they hate most; build agents for it. Ties to support deflection: teams now tackle interesting cases.
These require no risk hikes or process changes, unlike full autonomy attempts that fail.
Infrastructure Shifts for Agent Builders/Users
Apps now serve agent users, so prioritize APIs over UIs. Deployments must 'just run' for agent-written code—less emotional attachment to infra details.
Agent apps need sandboxes (e.g., author's 'just bash': TypeScript bash interpreter, zip-nanosecond startup). Security looms like 1999 web: separate harness execution from generated code runs (Anthropic's new agent does this; most don't). Expect paradigm shifts beyond general sandboxes.
Application Layer Wins as Models Commoditize
Build model-agnostic layers: innovation accrues to AI engineers, not labs. Europe leads (Vercel AI SDK: 10M downloads/week, Berlin-led; Pi coding agent from Austria; Open Claw). Two futures: labs monopolize (unlikely, AI stays expensive); or commoditization (cloud/GCP cheapest infra forces prices down)—favoring app-layer builders. Value in agents creating business outcomes, not base models.