Anthropic engineer's 14-minute guide to building Claude agents

My notes

Summary

A short TikTok recap of an Anthropic engineer’s 14-minute video on building agents in Claude. Three takeaways: start every agent with just three primitives (environment, one connected tool, system prompt), don’t agentify everything (use a decision tree to pick automated workflow vs. human-in-the-loop vs. autonomous agent), and respect the agent’s tiny working context (~10k-20k tokens) by trimming system prompts ruthlessly.

Key Insight

• Three-piece minimum viable agent: environment + one tool + system prompt. Anything more is premature scaling. Most agent failures come from stacking tools before the base loop works.
• Not every task is an agent. Decision tree: deterministic workflow with human approval > autonomous agent. Human-in-the-loop is the default; full autonomy is reserved for narrow, well-bounded tasks.
• Context window is the real constraint. Working budget for the agent’s reasoning is roughly 10k-20k tokens per turn after tool definitions, examples and system prompt are loaded. Bloated system prompts crowd out the actual task.
• Use Claude to design the agent itself. Feed Claude the takeaways and let it map the tool list, system prompt skeleton, and likely failure modes before any code is written. Cheaper than discovering them at runtime.
• This is a TikTok summary, not the source video, so treat the numeric claims (token budget, “three things”) as the engineer’s heuristics, not Anthropic guidance.