15 agent internals

after Hanako

15 things your agent does before it answers you see the last one

one turn
1/15

1. Renders the prompt

templateyou are {{role}}budget {{n}} turns
rendereda reviewer, 12 turns

template plus variables, not what you typed

2. Picks the context

api.tsspec.mdlog
context

which files and history get in

3. Trims the middle

184k tokens
head + tail kepttext-amber-700 dark:text-amber-400

what did not fit is simply gone

4. Loads tool schemas

schemas
6.2k tokens

every tool you defined costs tokens

5. Chooses one tool

read_filerun_testsopen_pr
one wins

one call, out of everything available

6. Fills the arguments

arguments
pathsrc/api.ts
moderw
guessed, not given

the path it guessed, not the one you meant

7. Waits on the tool

model 210ms
tool 2,840ms
queue 90ms

the latency nobody attributes correctly

8. Truncates the result

42k
900 tokens

42k of output, 900 tokens kept

9. Retries in silence

attempt 1timeoutattempt 2ok
never surfaced to you

the failure you never saw happen

10. Grades its own work

rubric score
passes at 8.0

only if you wrote a rubric

11. Decides to continue

check
continuestop

continue, retry, or stop

12. Compacts the history

one summary

your turn 3 becomes one sentence

13. Writes to disk

statestate.json
checkpointed

the only part that survives a crash

14. Spawns a subagent

mainsubagent
40k in, 1k back

burns 40k, hands back 1k

15. Picks what to show

reasoningfinal answertrace
youtext-orange-700 dark:text-orange-400

the one line you actually read