Ikibazo gihuza ibibazo · Hagati Urwego
Ibice hamwe nibibazo byo kwigunga
Gerageza impamvu abakozi batanga akazi kubantu bafite imiterere mishya, uko incamake ya handoff itakaza amakuru, kandi mugihe ibangikanye bifasha cyangwa bikomeretsa.
Ikibazo 1 Bya 8
Niki gitandukanya cyane subagent numukozi mukuru?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Niki gitandukanya cyane subagent numukozi mukuru?
- It uses a different programming language
- It always uses a larger model
- It works in its own fresh context window and sends back a condensed result
- It has no access to tools
- Ni ukubera iki kwuzuza idirishya rishobora kuba ikibazo na mbere yacyo ntarengwa?
- Quality often drops as irrelevant material piles up
- The model's weights change
- Tokens become cheaper
- Tools stop working
- Nuwuhe murimo ukwiranye na subagent?
- Choosing the overall architecture of a codebase
- Writing tightly linked parts of one function in parallel
- Keeping naming consistent across a project
- Searching a large repository for every caller of a function
- Nisoko nyamukuru yo gutakaza amakuru hamwe na subagents?
- Subagents delete files
- Summaries drop details, and subagents know only what the task description said
- Subagents cannot read text
- The main agent forgets its own instructions
- Kuberiki gushiraho abakozi benshi mubisanzwe bigura ibimenyetso byinshi?
- Each subagent reloads instructions and repeats some exploration
- Subagents are billed at double rates
- Summaries are longer than the original work
- Parallel work is always billed as sequential
- Why can parallel subagents produce incompatible results?
- They share one context window
- They always use different models
- They cannot see each other's work while running
- They run too slowly
- What should a good subagent task description include?
- Only a single word
- The main agent's full transcript
- Nothing, so the subagent stays unbiased
- The goal, constraints, scope and exact output format
- Why might a code-review subagent catch problems the main agent missed?
- It has more tools
- It has not seen the main agent's reasoning, so it is less likely to share the same assumptions
- It always uses a larger model
- It reads the code faster