5. Give Veltro a task
The tour was somebody else’s script. This step is yours.
Ask for something real
Section titled “Ask for something real”In Lucia’s conversation zone, or from a terminal with veltro '…', give it a
job that requires more than one tool. Something like:
read the files in /lib/veltro/demos and summarise what each one doesor
find every file under /appl that mentions cowfs and tell me how they relateVeltro will pick tools, run them, and show its working. What it picked is not a
mystery — it came from /tool/tools, which you read in step 4.
Watch what it actually did
Section titled “Watch what it actually did”While it works, or after it finishes, look at the trail:
; cat /tool/toolsEach tool directory has a doc file describing what it does and a schema file
giving its call signature:
; cat /tool/search/doc; cat /tool/search/schemaAn agent’s capability surface is a directory listing. You do not have to trust a description of what the agent can do — you can enumerate it.
Narrow it and try again
Section titled “Narrow it and try again”This is the part worth doing slowly, because it is the part that is different.
Look at what the agent has been granted:
; cat /tool/pathsThose are the paths it can reach. Not the paths it is allowed to reach — the paths that exist for it. Ask it to do something involving a path that is not on that list, and watch what comes back.
That failure is the subject of the next step, and it is worth arriving at it having tried this yourself first.
Verify
Section titled “Verify”You have this step when you have given the agent an instruction you invented, seen it complete, and can name which tools it used by reading the filesystem rather than by trusting its summary.
Ask it directly:
which tools did you just use, and what does each one do?Then check its answer against /tool/tools and the doc files. It should
match, and you should be able to prove it does.
Next: Namespaces contain — try to make the agent reach something it should not, and find out why it cannot.