Knowledge reuse across coding tasks
The Tenjin benchmark
The same coding tasks. The same agent. With and without Tenjin.
In an early Codex test
- Fewer tokens
- 8%
- Less agent time
- 7%
We gave Codex the same software engineering tasks, with and without access to prepared lessons in Tenjin.
A task from the benchmark
Fixing a crashing tool
An agent was assigned to repair a currency-formatting tool. It crashed because one JavaScript module loaded another incorrectly.
We gave fresh agents the same broken project. With Tenjin, they could retrieve a prepared note explaining that import mismatch. Without Tenjin, they had to work it out from the code.
Tenjin makes useful knowledge available to the next agent. This test measured reuse of prepared lessons; we tested the cost of capturing new lessons separately.
What we did
We ran each task three times per condition, using the same model and limits in a fresh environment.
Independent tests checked the finished work. We measured model tokens and the time the agent spent working, including its tools and Tenjin calls.
All comparisons and uncertainty
We also tested a Markdown file, Tenjin restricted to team lessons, and Tenjin capturing lessons during an earlier attempt. Lower ratios mean less usage; 1.00 is the no-Tenjin baseline. Brackets show 95% intervals. We compare tokens and agent time per verified completion, weighting each task equally.
- Without Tenjin
- Tokens: 1.000
- Agent time: 1.000
- 30 of 30 attempts passed.
- Shared Markdown
- Tokens: 0.856 [0.716–1.057]
- Agent time: 0.635 [0.453–0.863]
- 30 of 30 attempts passed.
- Tenjin with prepared lessons
- Tokens: 0.923 [0.791–1.092]
- Agent time: 0.934 [0.792–1.144]
- 29 of 30 attempts passed.
- Tenjin with team lessons only
- Tokens: 0.919 [0.778–1.071]
- Agent time: 0.798 [0.564–1.094]
- 29 of 30 attempts passed.
- Tenjin with capture and reuse
- Tokens: 1.397 [1.212–1.604]
- Agent time: 0.933 [0.693–1.204]
- 29 of 30 attempts passed.
The capture-and-reuse condition used about 40% more tokens at first reuse. The headline numbers measure access to prepared lessons, not the full capture-and-reuse process.
A separate Claude run used about 5% more tokens with prepared lessons, also inconclusive. We have not established a consistent saving across models.
Tasks and completion checks
The ten tasks covered test setup, import paths, JavaScript module compatibility, TypeScript execution, workspace setup, stale build files, and whitespace mismatches.
Across all five conditions, 150 attempts were scheduled: 147 passed and three were invalid. Every valid attempt passed. The three invalids were two setup refusals after Docker interruptions and one producer run that exhausted its subscription quota. None was replaced.
Codex CLI 0.154.0 · gpt-5.6-sol · low effort · September 13, 2026. Repeats measure variation on the same ten tasks; they are not 150 independent tasks.
What these results cover
This is a diagnostic experiment on small, controlled coding tasks. The headline intervals include no improvement, and invalid attempts left unequal valid coverage. It does not establish typical savings for a team.
Time covers the later agent’s work. Token accounting includes incremental capture at first reuse where applicable; the earlier task’s full effort is separate. Tokens are model usage, not a dollar bill.
The Markdown and Tenjin conditions did not have identical lesson sets, so that comparison does not isolate retrieval quality. Public fallback was enabled except in the team-only condition.
Try Tenjin with your team
Your repository has its own history. Start with a real task and see whether earlier knowledge makes a difference.