Your agent loop needs a budget, not a better prompt
Unbounded loops fail slowly and expensively. Treat steps, tokens, and time as product-level limits.
Featured pick
Your agent loop needs a budget, not a better prompt
Unbounded loops fail slowly and expensively. Treat steps, tokens, and time as product-level limits.
Latest
-
-
The trace is the artifact
If you cannot replay how an agent reached an answer, you cannot debug it, evaluate it, or improve it.
-
Tools are an interface, not an API
Models choose tools from descriptions, not type signatures. Design tool surfaces the way you design onboarding.
-
Context engineering for agents
Why the window is a budget to manage, not a bucket to fill, and the strategies that follow from that.
Why we recommend it The clearest treatment of the shift from prompt engineering to context management. The write/select/compress/isolate framing is useful immediately, even if you disagree with the specifics.
LangChain Blog
-
Agents
Chip Huyen's survey of agent architecture, planning, tool use, and the gaps between demos and production.
Why we recommend it The section on evaluation and failure modes is the part worth re-reading. It explains why agent demos look better than agent products, with the least hype of anything written that month.
Chip Huyen
-
Things we learned about LLMs in 2024
Simon Willison's year-in-review on model capabilities, costs, and what actually changed.
Why we recommend it The most useful annual snapshot of model progress, written for practitioners rather than investors. The cost and capability comparisons are the parts worth bookmarking.
Simon Willison
-
Building effective agents
Anthropic's taxonomy of agentic patterns, from prompt chaining to autonomous loops.
Why we recommend it Still the clearest map of the design space. It separates workflows from agents and gives each pattern a concrete failure mode, which most tooling marketing never does.
Anthropic Engineering
Scroll or use the arrow keys to browse.
All items
-
Your agent loop needs a budget, not a better prompt
Unbounded loops fail slowly and expensively. Treat steps, tokens, and time as product-level limits.
-
The trace is the artifact
If you cannot replay how an agent reached an answer, you cannot debug it, evaluate it, or improve it.
-
Tools are an interface, not an API
Models choose tools from descriptions, not type signatures. Design tool surfaces the way you design onboarding.
-
Context engineering for agents
Why the window is a budget to manage, not a bucket to fill, and the strategies that follow from that.
Why we recommend it The clearest treatment of the shift from prompt engineering to context management. The write/select/compress/isolate framing is useful immediately, even if you disagree with the specifics.
LangChain Blog
-
Agents
Chip Huyen's survey of agent architecture, planning, tool use, and the gaps between demos and production.
Why we recommend it The section on evaluation and failure modes is the part worth re-reading. It explains why agent demos look better than agent products, with the least hype of anything written that month.
Chip Huyen
-
Things we learned about LLMs in 2024
Simon Willison's year-in-review on model capabilities, costs, and what actually changed.
Why we recommend it The most useful annual snapshot of model progress, written for practitioners rather than investors. The cost and capability comparisons are the parts worth bookmarking.
Simon Willison
-
Building effective agents
Anthropic's taxonomy of agentic patterns, from prompt chaining to autonomous loops.
Why we recommend it Still the clearest map of the design space. It separates workflows from agents and gives each pattern a concrete failure mode, which most tooling marketing never does.
Anthropic Engineering