Tag: ai-agents

Define named subagents on agents in the Cursor SDK, let a code-reviewer delegate to a test-writer through the built-in Agent tool, and understand the real nesting depth cap: a subagent spawned by another subagent cannot spawn further ones, despite the changelog's 'and so on'.

By default a local Cursor SDK agent runs every tool call without asking. Set local.autoReview to route Shell, MCP, and Fetch calls through the classifier, then steer it with the autoRun block in permissions.json. With code, the three-step evaluation order, and why none of it is a security boundary.

For months availableModels restricted the model picker but left the Default option wide open. Claude Code 2.1.175 adds enforceAvailableModels so admins can finally pin a strict model allowlist.

The Cursor SDK already persists agent state to disk so Agent.resume() can pick a conversation back up after a crash. This walks the SqliteLocalAgentStore default, the JsonlLocalAgentStore you can commit to git, and how to implement the LocalAgentStore interface for Postgres, Redis, or an in-memory CI store.

Pass local.customTools to a Cursor SDK agent and the model calls your in-process functions through the built-in custom-user-tools MCP server. No stdio handshake, no separate process, no transport bugs. Local agents only: cloud agents throw ConfigurationError.

Use Quartz.NET when an LLM job must run on a real cron and never overlap itself, Hangfire when each run must survive a restart and retry on rate limits, and a plain BackgroundService only for a loose in-process loop. A decision matrix with the cron and concurrency gotchas that pick for you.

Rule-based checks are your floor and they are non-negotiable; LLM-as-judge is the ceiling you add when code quality, not just correctness, is what you ship. Here is the decision, with cost, latency, and the SWE-bench gap that proves why.

Claude subagents give you orchestration-level parallelism out of the box: one agent spawns isolated sub-contexts that run concurrently. OpenAI Assistants never had that, and it shuts down August 26, 2026, so new parallel work belongs on the Responses API plus your own fan-out. Here is how the two models actually differ and which to reach for.

Claude Code v2.1.169 (June 8, 2026) ships a --safe-mode flag that disables every customization for clean troubleshooting, and a /cd command that moves your session to a new directory without breaking the prompt cache mid-run.

The cache read and write multipliers are identical on both models, so the break-even point is the same. What differs is the minimum cacheable prefix (1,024 vs 4,096 tokens), the per-token dollar savings, and a new Opus 4.7 tokenizer that counts up to 35% more tokens. With claude-sonnet-4-6 and claude-opus-4-7 pricing math.