One structured agent is a reliable tool. But a batch of 200 performance reviews does not wait for you to run them one at a time. What's the gap between a single call and something that processes a whole quarter's worth of data?
Composing them. You'd want agents specialized for different steps — one that summarises, one that classifies — and then a loop that runs the right agent against each input in the list.
That's the whole week. Day 17 is two agents with different system prompts returning one dict. Day 18 is a pipeline that keeps every intermediate — summary and urgency together. Days 19 through 21 are batch patterns: a list comprehension calling an agent per item, min(outputs, key=len) to surface the sharpest response, and chained .output.split() inside a comprehension to get word counts across a whole inbox.
I'm chaining these like functions now. Summarise, classify, format — three agents in a list comprehension. One prompt is a tool. Many prompts is a product.
Exactly the shift. You stop thinking about a single call and start designing the pipeline first — inputs, agents, shape of output. By Friday that instinct is automatic.
system_prompt returning one dictmin(outputs, key=len) for the sharpest response.output.split() + len() across a batchGoal: you design multi-agent pipelines and batch patterns before writing a single line.
Create a free account to get started. Paid plans unlock all tracks.
One structured agent is a reliable tool. But a batch of 200 performance reviews does not wait for you to run them one at a time. What's the gap between a single call and something that processes a whole quarter's worth of data?
Composing them. You'd want agents specialized for different steps — one that summarises, one that classifies — and then a loop that runs the right agent against each input in the list.
That's the whole week. Day 17 is two agents with different system prompts returning one dict. Day 18 is a pipeline that keeps every intermediate — summary and urgency together. Days 19 through 21 are batch patterns: a list comprehension calling an agent per item, min(outputs, key=len) to surface the sharpest response, and chained .output.split() inside a comprehension to get word counts across a whole inbox.
I'm chaining these like functions now. Summarise, classify, format — three agents in a list comprehension. One prompt is a tool. Many prompts is a product.
Exactly the shift. You stop thinking about a single call and start designing the pipeline first — inputs, agents, shape of output. By Friday that instinct is automatic.
system_prompt returning one dictmin(outputs, key=len) for the sharpest response.output.split() + len() across a batchGoal: you design multi-agent pipelines and batch patterns before writing a single line.