Same engine, same rubric, same filter. What changes is who does the assessing, where the files live, and who starts a run.
Read the boxes as boundaries. Each one is a system somebody else operates, and every arrow crossing a boundary is candidate data leaving one place for another.
No cloud, no mailbox, no model configured, nothing billed. CVs go in a folder and a coding agent reads them.
flowchart TB
subgraph LOCAL["Your machine"]
direction LR
OPERATOR["you"]
OPEN["openings/*.yaml<br/>the rubric and the knockouts"]
CV["cvs/<br/>PDFs and text, dropped in by hand"]
CLI["talanton CLI"]
ASSESS["assessments/<br/>one YAML per candidate"]
end
subgraph ELSEWHERE["Wherever your coding agent runs - a model, not your machine"]
AGENT["your coding agent<br/>reads the CV, writes the assessment"]
end
OPERATOR --> |"1 - write the role"| OPEN
OPERATOR --> |"2 - drop the CVs in"| CV
OPEN --> CLI
CV --> CLI
CLI --> |"3 - next: the rubric, and the CV fenced as untrusted"| AGENT
AGENT --> |"4 - record: the assessment, unedited"| CLI
CLI --> |"5 - the knockout filter runs here, not in a prompt"| ASSESS
ASSESS --> |"6 - status / show"| OPERATOR
classDef actor fill:#ffffff,stroke:#9aa0a6,color:#202124
classDef data fill:#f1f3f4,stroke:#80868b,color:#202124
classDef svc fill:#e8f0fe,stroke:#1a73e8,color:#202124
class OPERATOR actor
class OPEN,CV,ASSESS data
class CLI,AGENT svc
style LOCAL fill:#f8f9fa,stroke:#dadce0,color:#5f6368
style ELSEWHERE fill:#f8f9fa,stroke:#dadce0,color:#5f6368
No Google service is involved, no credential is needed, and nothing here can
send to anybody. talanton itself makes no model call: there is no Vertex
project, no API key, and nothing is billed to you.
But be precise about the second boundary, because it is easy to read this mode as "the CV never leaves the room". It does. Your coding agent is a model, and step 3 hands it the CV text, so the file goes to whichever provider runs that agent under whatever terms you have with them. That is a smaller and more controllable exposure than a screening pipeline — one CV at a time, at your keystroke, to a provider you already chose — but it is not none, and a candidate would not consider it none.
Applications arrive by email. Files live on a shared drive so more than one person can read them. A model screens them. A person still starts every run, from their own machine.
flowchart LR
CAND(["candidate"])
OPERATOR(["operator"])
subgraph WS["Google Workspace"]
direction TB
APPLY["apply@ mailbox<br/>IMAP, read only"]
DRIVE["shared drive<br/>applications/ + assessments/"]
SEND["recruiting@ account<br/>SMTP, sends only"]
OPSBOX["operator's inbox"]
end
subgraph LOCALM["Your machine"]
direction TB
CLI["talanton CLI<br/>run as the service account"]
end
subgraph GCP["Google Cloud"]
direction TB
SM["Secret Manager<br/>mailbox app passwords"]
VERTEX["Vertex AI<br/>screener + root agent + correspondent"]
end
CAND -->|"1 - CV by email"| APPLY
APPLY -->|"2 - fetch, attachments only"| CLI
SM -.->|"passwords, at run time"| CLI
CLI -->|"3 - write the CV"| DRIVE
DRIVE -->|"4 - CV text"| CLI
CLI -->|"5 - assess"| VERTEX
VERTEX -->|"6 - the assessment"| CLI
CLI -->|"7 - write the assessment"| DRIVE
CLI -->|"8 - shortlist"| VERTEX
VERTEX -->|"9 - ids and links, never names"| SEND
SEND -->|"10 - the shortlist"| OPSBOX
OPSBOX --> OPERATOR
DRIVE -.->|"opens a CV, if allowed"| OPERATOR
classDef actor fill:#ffffff,stroke:#9aa0a6,color:#202124
classDef data fill:#f1f3f4,stroke:#80868b,color:#202124
classDef svc fill:#e8f0fe,stroke:#1a73e8,color:#202124
class CAND,OPERATOR actor
class APPLY,DRIVE,SEND,OPSBOX,SM data
class CLI,VERTEX svc
style WS fill:#f8f9fa,stroke:#dadce0,color:#5f6368
style GCP fill:#f8f9fa,stroke:#dadce0,color:#5f6368
style LOCALM fill:#f8f9fa,stroke:#dadce0,color:#5f6368
Three things the picture is meant to show:
- The two mailboxes are separate accounts. The side that reads applications cannot send, and the side that sends has no access to the mailbox. If the sending identity were ever misused, it could not read a single CV.
- Nothing reaches a candidate. The only outbound arrow ends at the operator's inbox. There is no tool that takes a candidate as a recipient.
- What is emailed carries ids and links, never names. Who may learn an identity is decided by who can open the file on the drive.
The same stages, moved off the laptop. The laptop is gone from the diagram, and that is the whole difference.
flowchart LR
CAND(["candidate"])
OPERATOR(["operator, reading mail"])
subgraph WS["Google Workspace"]
direction TB
APPLY["apply@ mailbox<br/>IMAP, read only"]
DRIVE["shared drive<br/>applications/ + assessments/"]
SEND["recruiting@ account<br/>SMTP, sends only"]
OPSBOX["operator's inbox"]
end
subgraph GCP["Google Cloud - no person involved"]
direction TB
SCHED["Cloud Scheduler<br/>fires on your cron"]
JOB["Cloud Run job: talanton cycle<br/>service account attached,<br/>so ADC is already the account"]
SM["Secret Manager"]
VERTEX["Vertex AI"]
end
CAND -->|"1 - CV by email"| APPLY
SCHED -->|"2 - start a run"| JOB
APPLY -->|"3 - fetch"| JOB
SM -.->|"passwords"| JOB
JOB -->|"4 - write the CV"| DRIVE
DRIVE -->|"5 - CV text"| JOB
JOB -->|"6 - assess, then shortlist"| VERTEX
VERTEX -->|"7 - the assessment"| JOB
JOB -->|"8 - write the assessment"| DRIVE
VERTEX -->|"9 - ids and links"| SEND
SEND -->|"10 - the shortlist"| OPSBOX
OPSBOX --> OPERATOR
DRIVE -.->|"opens a CV, if allowed"| OPERATOR
classDef actor fill:#ffffff,stroke:#9aa0a6,color:#202124
classDef data fill:#f1f3f4,stroke:#80868b,color:#202124
classDef svc fill:#e8f0fe,stroke:#1a73e8,color:#202124
class CAND,OPERATOR actor
class APPLY,DRIVE,SEND,OPSBOX,SM,SCHED data
class JOB,VERTEX svc
style WS fill:#f8f9fa,stroke:#dadce0,color:#5f6368
style GCP fill:#f8f9fa,stroke:#dadce0,color:#5f6368
An assessment-only variant deploys assessor_app instead. It holds no tool
that can reach anybody, so nothing is emailed at all and a person reads the
assessments on the drive. That is the right thing to run unattended.
All of it is in talanton.toml, plus who you are when you run it.
| Mode 1 | Mode 2 | |
|---|---|---|
[storage.cvs], [storage.assessments] |
backend = "local" |
backend = "gdrive" and a folder_id |
[screening] |
absent | project, location = "global", model |
[mail.inbound] |
absent | user, password_secret |
[mail.outbound] |
absent | user, password_secret, operators, allow_domains |
| identity | your own login | the service account, impersonated |
| what assesses | your coding agent | the screener, on Vertex |
| commands | next, record |
fetch, assess, shortlist, cycle |
Two of those rows are the ones people get wrong.
location is global. No Gemini model is served from any europe-west
region: a regional value returns 404 rather than falling back. Storage can
still be pinned to Zurich, screening cannot, and the gap between those two is a
data-protection decision worth making deliberately rather than discovering.
Identity is not a detail. Run as the service account, not as yourself:
gcloud auth application-default login --impersonate-service-account=$SAGoogle Drive's drive.file scope reaches only files that identity created, so
a CV fetched under a personal login is one that only that person's talanton can
read. Everybody else gets an empty pipeline and no error. See Whose identity
talanton runs as in AGENTS.md.
Leave dry_run = true until somebody has watched a full run and read what it
would have sent. talanton check says loudly when it is off.
The table above reads like a ladder. It is not: storage backend, mailbox, who assesses and who triggers are four independent settings, and the code treats them that way. Any combination is valid.
The one worth naming is shared storage with no model: backend = "gdrive"
with [screening] absent. Real Drive folders, so more than one person can read
a CV and access is a membership list — and a coding agent doing the assessing,
so there is no Vertex project, no model id, and nothing billed. For a small
team that is often the most useful cell in the grid.
| value | |
|---|---|
[storage.cvs], [storage.assessments] |
backend = "gdrive" and a folder_id |
[screening] |
absent |
[mail.inbound] |
optional — fetch works with no model |
[mail.outbound] |
optional — needed only for send |
| what assesses | your coding agent, through next and record |
The loop is fetch → next → read it → record → status → send. Nothing
is weakened by taking it: record runs the same knockout filter as the agent
path, send runs the same name check, and talanton check passes on this
shape — a missing model is reported as a note, not a failure.
Two things carry over that people miss.
Identity still decides what you can see. Everything under Identity is not
a detail above applies here, model or no model. drive.file reaches only the
files the calling identity created, so fetching as the service account and then
running next as yourself gives "nothing waiting" and no error.
Mode 1's privacy note does not carry over. CVs are on Drive now, which is the point of choosing this cell. And the part about your coding agent being a model applies exactly as it does in mode 1.
Same config file. What changes is where the process runs and how it gets its identity:
| Mode 2 | Mode 3 | |
|---|---|---|
| where it runs | your laptop | Cloud Run, GCE, GKE, or Agent Engine |
| how it gets its identity | you impersonate the service account | the account is attached, so ADC is already it |
| logging in | again whenever the token expires | never |
| what starts a run | you type a command | Cloud Scheduler, or cycle --watch on IMAP IDLE |
This is why the identity work matters more than it looks. A deployment resting on a human token cannot run unattended, because recovering that token needs a person, a browser and a password.
# the whole agent, including the send path
adk deploy agent_engine --project=$GOOGLE_CLOUD_PROJECT \
--region=$GOOGLE_CLOUD_LOCATION --display_name="Talanton" talantonThe stages also deploy separately, each with only the privileges it needs:
fetch needs the mailbox and cannot send, assess needs the two locations and
Vertex, shortlist needs the sending account. Deploying assessor_app gives
you the first two with no send path anywhere in the process.