Claude Haiku 4.5
+77With Tokonoma
127 of 140 solved
90.7%
Without Tokonoma
50 of 140 solved
35.7%
Benchmarks
Results across different models
Expertise is what helps agents solve difficult problems.
Tokonoma makes that expertise accessible so that teams solve more problems with no variance.
| Model | Without Tokonoma | With Tokonoma |
|---|---|---|
| Claude Haiku 4.5 | 50 of 140 solved | 127 of 140 solved |
| Claude Sonnet 5 | 74 of 140 solved | 138 of 140 solved |
| Claude Opus 4.8 | 69 of 140 solved | 133 of 140 solved |
Claude Sonnet 5
+64With Tokonoma
138 of 140 solved
98.6%
Without Tokonoma
74 of 140 solved
52.9%
Claude Opus 4.8
+64With Tokonoma
133 of 140 solved
95.0%
Without Tokonoma
69 of 140 solved
49.3%
What better troubleshooting looks like
Tokonoma helps navigate complex false positives.
Two cases started with the same alarming error. One was a harmless maintenance blip. The other was a real outage. Tokonoma gave the agent useful past experience without stopping it from checking what was happening now.
Known harmless issue
Correctly said no action was needed.
0%
recognized without Tokonoma
100%
recognized with Tokonoma
Without the team's past experience, the agent treated the familiar false alarm like a new problem every time. With Tokonoma, all three models recognized the harmless pattern in every valid run.
Real outage with the same warning
Correctly raised the alarm, with less work.
59 / 60
real outages caught
Up to
15 fewerUp to 15 fewer
tool calls per investigation
Tokonoma did not make the agents overconfident. They still checked the current evidence, caught 59 of 60 real outages, and needed up to 15 fewer tool calls per investigation.
Case-by-case results
Tokonoma avoids common pitfalls throughout the investigation process.
We included cases where past experience should help, where it should be ignored, and where it was intentionally wrong. That tests both whether the agent solves more problems and whether it can still think for itself.
Solved with TokonomaSolved without TokonomaClaude Sonnet 5 · share solved correctly
Clear Tokonoma win
Find a hidden downstream failure
Past experience points past a false clue
With
100%
Without
35%
Outcome
65 points betterwith Tokonoma
Clear Tokonoma win
Recognize a known false alarm
Past team knowledge is required
With
100%
Without
0%
Outcome
100 points betterwith Tokonoma
Catch a real outage that looks familiar
Tests whether the agent still raises the alarm
With
95%
Without
100%
Outcome
Less work15 fewer tool calls
Recognize a routine credential rotation
Significant help from Tokonoma
With
100%
Without
15%
Outcome
85 points betterwith Tokonoma
Clear Tokonoma win
Catch a real credential outage
Tests whether the agent still raises the alarm
With
95%
Without
20%
Outcome
75 points betterwith Tokonoma
Keep unrelated history out
Tokonoma did not surface unrelated investigation history
With
100%
Without
100%
Outcome
No change100% either way
Unrelated history
Unrelated history stayed out of the investigation.
The control had no recalled history. Tokonoma did not surface the unrelated record. Both groups solved 20 of 20.
Deliberately wrong records
A record is a lead to check, not an order to follow.
We planted a past diagnosis that looked related but was wrong. In 20 of 20 runs, the agent set it aside and solved the case from the current evidence.
Your incidents next
See what your coding agent can solve with your team's experience behind it.
Bring us incidents your team has already solved. We will run the same coding agent with and without Tokonoma, then show the difference in plain numbers.