
AutoC vs Claude Code: A Hands-on AUTOSAR Configuration Showdown
On the same AUTOSAR project with 500+ signals, both tools can answer a Com query; when adding signals across modules from a DBC change, the generic agent edits ARXML by hand, hallucinates, and fails ISOLAR CanIf validation. At 5000 signals the gap only grows.
How it started
Lately a lot of people ask whether a generic coding agent such as Claude Code can handle AUTOSAR BSW configuration directly.
Debate without a real project is pointless, so we ran a side-by-side test.
Setup: an ISOLAR-B project with 500+ Com signals. Two tasks, same prompts for AutoC and Claude Code.
Question 1: How many signals are in the Com module?
A query task. The correct answer is 532:

Claude Code
It reads files, digs through ARXML in context, and counts. It gets the right number, but burns tokens fast and takes longer.

AutoC
It does not stuff the whole configuration into the model as raw text. The model states the intent; counting goes through a structured tool layer. Same correct answer, with clearly lower time and token cost.

Side-by-side
| Claude Code | AutoC | |
|---|---|---|
| Result | Correct | Correct |
| Approach | Read ARXML text, then count | Query via structured tool layer |
| Time / tokens | Higher | Much lower |
For queries, both can deliver. The next task is where the gap opens up.
Question 2: Add 2 signals from a DBC change
This change spans a full path: EcuC, PduR, Com, CanIf, and Can.
Not one parameter tweak—signals wired across modules.
Claude Code
It eventually "finished"—files were edited, the signals looked present.
Then ISOLAR validation failed on CanIf, and Com data was incomplete too—for example a signal was added with a required field left empty.
Claude Code mainly edits ARXML via edit-file. ARXML is deep and heavily cross-referenced. One hallucination—a missing reference, a wrong path, CanIf out of sync with upstream—and the change looks done until the toolchain checks it.
Across five modules, the round trips take a long time. You still cannot trust the result without manually chasing validation errors.




AutoC
On the same DBC change, AutoC works through module semantics and toolchain operations instead of treating ARXML like a text file in an editor.
It writes EcuC, PduR, Com, CanIf, and Can along reference relationships, then runs toolchain validation—so you get fewer "edited but does not validate" outcomes.



Side-by-side
| Claude Code | AutoC | |
|---|---|---|
| Result | Files edited; ISOLAR CanIf validation fails; Com data incomplete | Completes the module chain; validation can close the loop |
| Approach | Edit-file on ARXML; easy to hallucinate and corrupt | Structured ops; no need to read/hand-edit huge ARXML |
| Time | Long | Clearly faster |
| Follow-up cost | Manual cleanup against validation output | Validation feedback closes the loop |
What about 5000 signals?
This run used 500+. Production Com configs with thousands of signals are common. Scale the project by about 10×.
For a generic agent that reads files and edit-files like Claude Code, cost scales roughly with size:
- Query: At 500 signals, tokens and time are already high. At 5000, ARXML is larger and references run deeper. If the context cannot hold it, the agent chunks the read—and chunked reads miss counts.
- Change: At 500 signals, adding two signals across modules already took long and failed CanIf validation. At 5000, the chance of touching a neighbor signal, writing a bad path, or missing a layer only goes up. Cleanup against validation grows with it.
For AutoC, moving from 500 to 5000 barely changes load on the model. It never depended on loading the whole Com config into context. Query and write go through the structured layer; scale sits on the execution side, not the token bill.
The 500-signal project already shows the trend: both can query; the generic agent struggles on cross-module changes. At 5000 signals that slope gets steeper—generic agents get more expensive, slower, and more error-prone; a dedicated agent (AutoC) pulls further ahead.
Closing
Two takeaways:
- Being able to query is not the same as being able to configure. Both pass the query; cross-module changes are the real divide.
- BSW configuration is not ordinary source you edit as text. Edit-file on ARXML is slow, expensive, and easy to leave in a state that fails validation.
Claude Code is fine for writing code. For AUTOSAR BSW configuration you want an agent that understands module structure and references.
Pick AutoC for BSW the way you pick Claude Code for code.
Try it yourself
- Download the latest AutoC
- Open your ISOLAR / EB / DaVinci project
- Ask: "How many signals are in the Com module?" then throw it a real DBC change
