- Reads the ticket and the repository, nothing else.
- 0 of 11 off-repo team rules followed
- Hidden tests fail: 0 of 10 runs fully correct
Code that is right on the first try.
With Mios, your agents follow your team’s decisions, even the ones that are not in the code. They get it right the first time far more often, need half the follow-ups, finish twice as fast, and your Claude bill goes down.
- Right the first time5 / 10first answers fully correct with Mios, vs. 0 of 10 for Claude Code alone
- Back-and-forths−52%prompts to reach correct code, vs. Claude Code alone
- Time−49%time to reach correct code, vs. Claude Code alone
- Claude bill−20%on your Claude Code bill until the code is right, compared with Claude Code alone
Runs from November 12 to 14, 2026 · Mios build of November 14 · Claude Sonnet 5.5
The rule is not in the repository.
A ticket asks for a feature across route, controller, service, test and docs. The rule it must follow was written by the team in chat.
- Mios hands it the team’s rules before it writes.
- 10 of 11 off-repo team rules followed on average
- Every hidden test green in 5 of 10 runs
Right on the first try.
One prompt, just the ticket. With Mios, 5 of 10 runs pass every hidden test; with Claude Code alone, 0 of 10. No other setup gets a single one.
- Claude Code alone
- + Obsidian
- + Graphify
- + Obsidian and Graphify
- + Mios
- Claude Code alone
- + Obsidian
- + Graphify
- + Obsidian and Graphify
- + Mios
Fewer back-and-forths.
Every setup ends with all tests passing. With Mios it takes 52% fewer prompts and 49% less time than Claude Code alone.
- Claude Code alone
- + Obsidian
- + Graphify
- + Obsidian and Graphify
- + Mios
- Claude Code alone
- + Obsidian
- + Graphify
- + Obsidian and Graphify
- + Mios
Your Claude bill goes down.
−20% on your Claude Code bill with Mios, five tasks combined.
- Claude Code alone
- + Obsidian
- + Graphify
- + Obsidian and Graphify
- + Mios
Measured on five tasks, two runs each.
What it changes for your team.
- 10 / 11
off-repo team rules followed on the first answer, vs. 0 for Claude Code alone
- 90%
of hidden checks passed on the first try, vs. 23% for Claude Code alone
- 8
prompts to finish the five tasks, vs. 16.5 for Claude Code alone
How we measured it.
Five real feature tasks on a SaaS product in production.
11 decisions made in team discussions, never written in the repo. Mios captures them.
Hidden tests grade every answer. No model does the grading.
Five setups, one model: Claude Sonnet 5.5, same clone, same ticket.
One identical review comment for every setup, until all tests pass.
Detailed method
- 01Five real feature tasks
Five features in the repository of a SaaS product in production. Each one touches several files and plugs into existing code: route, controller, service, test and docs. The agent gets the ticket and nothing else.
- 02Hidden tests
The agent never sees the tests that grade it; they are copied in only for grading. One check per task for the feature itself, one per team rule, 24 checks in all. No model takes part in the grading.
- 03Rules written outside the code
Every task depends on rules the team wrote in chat, shared documents and messaging. 11 of them appear nowhere in the repository, so an agent follows them only if something brings them in.
- 04Five setups, one model
Claude Code alone, with Obsidian on the repository’s docs, with Graphify on the repository, with both, and with Mios. Every setup runs Claude Sonnet 5.5 on the same clone, with the same ticket and the same time limit.
- 05Two runs per cell
Two runs per task and setup. On the one task where the two runs disagreed, we added a third run for Mios and for the next best setup.
- 06The same review for everyone
When the first answer fails a test, the same session gets a review comment on the failing parts, word for word the same for every setup. From the fifth prompt on, it also states the missing rule and shows the failing tests, so every setup ends with correct code. Up to 8 prompts.
- 07Criteria set in advance
What counts as a win and what counts as noise was written down before the first run. Two things changed along the way: the prompt cap went from 4 to 8, and the review loop is now counted per task instead of per run.
All the numbersboth tables, with costs and tokens
First answer, by setup
| Setup | Fully correct | Checks passed | Off-repo rules followed |
|---|---|---|---|
| Claude Code alonethe ticket and the repository | 0 of 10 runs | 23%5.5 of 24 | 0 of 11 |
| + Obsidianthe repository’s docs as notes | 0 of 10 runs | 40%9.5 of 24 | 0.5 of 11 |
| + Graphifya graph of the repository | 0 of 10 runs | 33%8 of 24 | 0 of 11 |
| + Obsidian and Graphifyboth of the above | 0 of 10 runs | 33%8 of 24 | 0 of 11 |
| + Miosthe team’s memory over MCP | 5 of 10 runs | 90%21.5 of 24 | 10 of 11 |
Off-repo rules followed: the average over the runs, out of the 11 rules found nowhere in the repository.
To correct code, five tasks combined, by setup
| Setup | Prompts | Time | Claude Code tokens | Claude Code cost |
|---|---|---|---|---|
| Claude Code alonethe ticket and the repository | 16.5reference | 954 sreference | reference | $2.241reference |
| + Obsidianthe repository’s docs as notes | 17.5+6% | 1,029 s+8% | +10% | $2.352+5% |
| + Graphifya graph of the repository | 17+3% | 1,064 s+12% | +9% | $2.330+4% |
| + Obsidian and Graphifyboth of the above | 17+3% | 1,055 s+11% | +17% | $2.423+8% |
| + Miosthe team’s memory over MCP | 8−52% | 487 s−49% | −33% | $1.796−20% |
Claude Code tokens and cost are read from each session. Per task, the median says the same: 1.5 prompts for Mios vs. 3 for Claude Code alone, 87 s vs. 184 s.
Give your agents your team’s rules.
Mios is in private beta and opens to a few teams at a time.
Mios opens to a few teams first. Book a call and we’ll get yours in.