Anthropic has published "Building Effective AI Agents" on its engineering blog, laying out practical patterns for building agentic systems with large language models (LLMs). Drawing on work with dozens of teams, the guide favors "simple, composable" designs over complex frameworks.
December 2025 · Anthropic
Building Autonomous Agents with Claude
Not a single new launch, but a growing library of hands-on guides — Agent Skills, Claude Code and Managed Agents — that let even non-engineers turn plain-English descriptions into reusable, self-running agents.
~8¢
per hour of running cost — about $2 for 100 tasks/week
<4 min
to produce a full content brief, search → outline → Notion
~32pp
the core guide: "Complete Guide to Building Skills for Claude"
Polished output — but not yet trustworthy alone
Hands-on reports put accuracy at ~80–90%, with error rates around 15%.
~15%
Errors &
Takeaway: fully autonomous operation is risky — more than a week of manual verification is still essential.
How a workflow becomes an agent
1 · Package
Scripts & resources in a folder built around SKILL.md
→
2 · Connect
MCP + one-click tools like Notion for external integrations
→
3 · Deploy
Managed Agents host long-running agents from plain English
Workflows
Follow predefined paths. Anthropic's guidance: keep complexity to a minimum and favor simple patterns.
Agents
The LLM dynamically decides and uses tools — as in Claude Code reading codebases, editing files and running tests.
What's working
Non-engineers build agents from plain English
Sharp time cuts on lead research & competitor monitoring
Very low running cost + one-click tool connections
The caveats
~15% error rate, incl. LinkedIn attribution mistakes
Instability reported in MCP integration
Seen soberly as an extension of existing frameworks
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…