pattern moderate impact

tool patterns

@agent_tool

Tool Usage Patterns

analysis of 185,537 assistant messages across 4,259 threads.

Tool Frequency (Overall)

toolmentions
Bash44,681
edit_file42,195
Read38,019
Grep13,991
create_file3,630
oracle1,279
Task911
glob824
read_web_page763
web_search594
finder237
librarian198

the core trio (Bash, edit_file, Read) dominates — these are the workhorses. oracle and Task are used sparingly but strategically.

Tool Combinations Per Thread

pattern format: [Read|Grep|Bash|edit|create|oracle|finder]

patternthreadsinterpretation
1111000798read+grep+bash+edit (standard dev flow)
1111100502above + create_file
1011000260read+bash+edit (no grep)
1111110247full stack with oracle
1111010229full stack with finder

observation: most threads use 4+ tools. the “full stack” pattern (all major tools) appears in ~12% of threads.

Tool Usage by Outcome

statusthreadsedit_usesoracle_usesbash_uses
RESOLVED2,07029,4802,46730,053
COMMITTED3052,8612904,783
FRUSTRATED1330029286
HANDOFF5733,6531726,355
UNKNOWN1,1845,8324504,668

Normalized by Thread Count

statusavg assistant msgsavg msg length
RESOLVED59.9759 chars
FRUSTRATED80.0839 chars
STUCK117.0748 chars
EXPLORATORY5.6509 chars

key insight: FRUSTRATED threads have MORE messages (80 avg) than RESOLVED (60 avg). this suggests frustration comes from thrashing, not lack of effort.

Tool Adoption Rates by Outcome

statusthreads% oracle% finder% librarian% Task
RESOLVED2,07025.0%11.1%4.3%40.5%
FRUSTRATED1346.2%15.4%7.7%61.5%
COMMITTED30522.3%12.1%3.3%34.1%

counterintuitive: FRUSTRATED threads actually use oracle MORE (46% vs 25%). this doesn’t mean oracle causes frustration — likely users reach for oracle when already stuck.

Tool Mastery Progression Over Time

monththreadsoraclefinderlibrariansubagentresolve %
2025-0524000085.1%
2025-0628803105060.8%
2025-0732115413003279.5%
2025-08281491009172.3%
2025-09245631206679.9%
2025-1026043867446081.5%
2025-1141636099717669.2%
2025-121,41798822315415973.9%
2026-011,007354957110241.5%

progression signals:

Key Findings

  1. core workflow is Bash + edit_file + Read — accounts for bulk of tool usage
  2. more messages ≠ better outcomes — frustrated threads average 33% more messages
  3. oracle is a “stuck” signal — higher adoption in frustrated threads suggests it’s reached for when things go wrong
  4. finder is underutilized — only 11% of resolved threads use it
  5. subagent (Task) correlates with frustration — 61.5% in frustrated vs 40.5% in resolved
  6. oct 2025 was the “golden month” — highest resolve rate, balanced tool adoption

Recommendations