How an agent finds its way around
Ask a coding agent to fix a failing test and it does what a new hire would do: it searches, opens the files that look related, follows imports, and reads until it is confident it has the picture. Then it writes the change.
Context takes room, even when caching changes how it is billed. Repeated reads can make a session harder to navigate, and additional turns can add usage. Whether that becomes an extra bill depends on your plan, billing method, and provider.
Where the waste comes from
Most of the extra reading falls into a few patterns:
- Whole files opened to find one function.
- The same file read again after the agent loses track of it.
- Exploratory reads of code that turns out to be unrelated.
- Retries that repeat a command that already failed, along with the context around it.
None of this is the agent being careless. Without a map of the repository, reading is the only way to learn it. The problem is that you pay for the learning on every task.
What changes with Mikie
Mikie gives the agent the map up front. When you send a request, Mikie adds a short note to it: the code the request is about and the tests that already use it, with file paths and line numbers. It asks the agent to read exact line ranges, run the tests once and stop when they pass, and it stops wasteful reads, such as a whole generated file, before they fill the conversation.
The agent still decides what to do. It simply starts from the right few places instead of discovering them one file at a time.
What we measured
In a held-out benchmark of 36 tasks in Python, TypeScript and Rust (2026-10-07), Codex worked the same tasks with and without Mikie, and every fix was accepted both ways. With Mikie, total tokens per run fell from 23.67M to 6.32M, a 73.3% reduction; cost fell 60.4% at list prices, and the agent made 70.6% fewer tool calls.
Those are tasks from four open-source repositories, run with one agent and one model on one workstation, on Mikie 1.1.131. Your repository can behave differently. Shadow mode can identify work worth inspecting; accepted active runs and usage records establish what changed in practice.
Give the agent a starting point it can verify
A useful task note contains the failure, the likely entry point, and the acceptance command. It leaves room to follow a dependency when the first explanation is wrong.
Fix the whitespace parsing failure.
Start with parse_duration in src/duration.py and its tests.
Read the relevant function and callers before editing.
Preserve the public API. Add the regression, then run the focused test.
Widen the search if the failure points outside this module.This is an illustrative prompt. Replace its paths and behavior with the failure in your repository; a precise starting point is useful only when it is correct.
Know when a short read is not enough
If the same function is reopened, write a short note with the current facts, changed files, and next test. If imports or a failing integration test contradict the local explanation, inspect that dependency before trimming more context.
- Keep acceptance checks and permission rules visible.
- Exclude generated output only when it is unrelated to the defect.
- Compare accepted work and retries alongside usage; a smaller context that misses the bug is a worse result.
Sources: Figures are from Mikie's public evidence record for the 36-task held-out benchmark of October 7, 2026. Tokens are provider-reported, cost is calculated at list prices, and session times are not published.