Schedule tasks on the web
Anthropic’s new “scheduled tasks” for Claude Code, which let AI agents run recurring jobs against cloud-hosted code and tools, is being treated as a step toward fully automated software workflows—from triaging issues and scanning logs to opening pull requests overnight. Commenters are split between enthusiasm for the productivity boost and worries about high inference costs, reliability, security, and vendor lock‑in, with several arguing that many proposed use cases are better handled by traditional, deterministic automation like cron plus scripts. The feature also feeds into a larger debate over how far current LLMs can be trusted to own end‑to‑end development loops, and whether centralizing memory and workflows with model providers will erode user control over their own tooling.
Pricing, Limits, and Throttling
- Confusion over pricing: some assume per-token; others clarify it’s subscription-based.
- Users report limits like “3 daily cloud scheduled sessions” even on higher tiers.
- Concerns about recent/likely throttling and changing usage terms, sometimes announced informally (e.g., via social media), feeding distrust.
- Some see these limits as rational congestion pricing; others view them as “rug pulls” and erosion of trust.
Future of Agentic Software Development
- Many describe a near-future loop: user feedback → AI-curated ticket → AI PR → AI review → deployment, plus A/B tests, telemetry, and progressive rollout.
- Some are already auto-generating PRs from GitHub issues and reviewing in ephemeral environments.
- Others strongly doubt end-to-end automation for anything beyond simple CRUD/web tasks, especially in safety‑ or finance‑critical domains.
Quality, Reliability, and Maintainability
- Repeated reports that current models:
- Handle small, localized changes well.
- Struggle with complex, unfamiliar, or long‑lived systems.
- Introduce tech debt: duplicated logic, inconsistent patterns, performance issues.
- Several engineers say careful AI‑assisted coding is faster than fully autonomous agents, because oversight is still essential.
- Debate over whether scaling laws and RL will inevitably push coding agents to superhuman performance vs. hitting limits in reasoning, context, and learning.
Inference Cost, Environment, and Energy
- Inference viewed as the main economic bottleneck; calls for cheaper, more efficient hardware and software.
- Some argue lower costs will just increase usage (more tokens, more agents); others highlight GPU capacity and environmental concerns.
- Environmental impact is contested: some say AI’s footprint is overstated; others insist training and energy use must be factored in.
Cron vs. “Scheduled Tasks on the Web”
- Many note this is essentially “cron + Claude in the cloud.”
- Proponents: useful for non‑devs, removes need for local infra, integrates with MCP tools (Slack, Sentry, GitHub, etc.).
- Critics: trivial to replicate with cron + API, potentially expensive, and introduces platform lock‑in and GitHub‑only assumptions.
Example Use Cases
- Scheduled security/package audits, Sentry triage, and code review reports.
- Auto‑triaging GitHub issues, generating PRs, and updating documentation.
- Some want richer capabilities (screenshots, arbitrary HTTP, self‑hosted repos) and turn to alternative “AI cronbox” services.
Security and MCP / Tooling Concerns
- MCP seen as powerful but also part of recent security incidents; debate over whether it’s inherently risky or just “an API.”
- Prompt injection framed as analogous to social engineering: no complete technical fix, only mitigations with trade‑offs.
Vendor Lock‑In and Control
- Strong current of worry about model providers owning memory, workflows, and tooling.
- Some advocate keeping agents and automation outside proprietary ecosystems, treating models as swappable commodities.
- Others welcome the convenience of integrated stacks and are less concerned about centralization.