Agents, autonomy, and collapsing app prices
Replit is betting the future of software on autonomous agents and the infrastructure that lets them thrive. CEO Amjad Masad argues that as anyone can generate complex apps with a prompt, traditional application software margins will erode toward zero, shifting value to problem-solving, orchestration, and agent “habitats.” The near-term roadmap centers on end-to-end testing, parallel simulations, and transactional environments to push agents from minutes of autonomy to hours—and eventually to reliable, large-scale autonomy.
Points clés
- Replit has shifted “all resources into agents” since late 2023/early 2024 to make software creation accessible to anyone without coding.
- Software engineering performance on the SWE-bench benchmark is reportedly at 70–80%, signaling rapid automation of large portions of coding work.
- Replit emphasizes the “habitat” agents need: cloud VMs, sandboxing, full Linux shell access, package install (system and language), multi-language support, deployments, databases, and scalability to millions of users.
- Built-in services already handled by Replit include one-line authentication, secrets management, background jobs, storage, domain linking, and turnkey deployments—because “coding is the easy part,” the surrounding infrastructure is the hard part.
- Upcoming roadmap features include universal model access (Replit handles billing/API integration) and payments, including agent wallets to buy tools (e.g., Twilio) or hire humans (e.g., TaskRabbit) and other agents.
- Replit frames agent autonomy in levels: language servers (L1), code completion/Copilot (L2), Replit Agent v2 at 3–3.5 (10–15 minutes autonomy), and v3 targeting L4 with longer, semi-supervised runs.
- Agent v3 rests on three pillars: end-to-end app testing via “computer use,” test-time compute with reversible, transactional file systems enabling cheap forks and parallel solution sampling, and auto-generating fast, reliable tests for every feature.
- Strategic prediction: application software prices trend toward zero “on the order of years,” as any user can generate bespoke apps with a prompt; value shifts from apps to solving problems directly.
- Evidence of disruption today: a Replit HR teammate built a bespoke org chart app in three days that replaces market offerings costing “tens of thousands of dollars per year.”
- Longer-term technical bets include multi-agent ecosystems (MCP seen as traditional RPC, not true agent-to-agent), and AlphaZero-style RL/self-play to avoid model collapse when code is agent-generated, alongside Replit’s deep use of NixOS and transactional environments.
À retenir
Start treating agents like interns who never sleep: give them guardrails, clear tests, and small budgets, then let them iterate. Learn enough prompting and product sense to be dangerous, capture your domain quirks so your agent doesn’t reinvent the wheel, and prototype end-to-end—auth, storage, deploy—so you’re solving problems, not just shipping “yet another app.” And yes, prices are heading south; your margins won’t save you, but your ability to orchestrate agents, run parallel experiments, and deliver outcomes just might—so sharpen those generalist muscles before the robots ask for PTO.
Sources
Quiz sur la vidéo: 5 questions
