THE HUMAN LAYER FOR AGENT WORK
Human direction.
Agent execution.
Many agents. One point of view.
A place to see what happened, understand why,
and decide what happens next.
01 / THE REAL DESIGN PROBLEM
Work is multiplying.
Attention isn’t.
A code agent opens a pull request. A research assistant leaves a recommendation. Another run is blocked on a billing rule. The work is moving. The person responsible needs a clear account of it.
Vantage explores that moment through Northbank, a fictional software team managing payment retries and usage billing. It brings the task, source, cost, evidence and decision into one review context.
Seven tabs. One decision.
The diff is in one place, the instruction in another, and the test failure in a third. Review begins by rebuilding context.
“Done” needs evidence.
A retry patch can pass a unit test and still retry a cancelled mandate. Passing checks and sound policy are separate questions.
Who allowed the action?
Scope, spend and permissions need an owner. Every approval should leave a reason that another person can inspect.
02 / THE PRODUCT
Your morning.
In full view.
Open a run. Inspect its evidence. Make a decision.
The queue stays in place while you review.
Changes
Review agent work across repositories before it lands.
| Run | Agent | Files | Changes | Trust | Status | Updated | Open | |
|---|---|---|---|---|---|---|---|---|
| NB-4821 billing-core | 3 | +58 −10 | L3 | Awaiting decision | 4m | |||
| NB-4817 billing-core | 4 | +96 −41 | L3 | Blocked | 1h | |||
| NB-4824 research | 1 | +34 −0 | L2 | Awaiting decision | 26m | |||
| NB-4822 billing-core | 2 | +41 −9 | L3 | Running | 11m | |||
| NB-4819 research | 1 | +28 −0 | L2 | Awaiting decision | 5h | |||
| NB-4814 merchant-dashboard | 2 | +52 −3 | L3 | Awaiting decision | 9h | |||
| NB-4825 comms | 1 | +18 −6 | L1 | Queued | 2h | |||
| NB-4816 analytics | 1 | +22 −4 | L2 | Landed | 3h | |||
| NB-4811 billing-core | 2 | +47 −16 | L3 | Landed | 8h | |||
| NB-4809 research | 1 | +31 −0 | L2 | Landed | 1d | |||
| NB-4806 comms | 1 | +9 −7 | L1 | Rejected | 1d | |||
| NB-4803 merchant-dashboard | 1 | +26 −8 | L3 | Running | 1d |
SAMPLE DATA · LOCAL DEMONSTRATION · NO EXTERNAL ACTIONS
03 / HUMAN JUDGMENT
The final say
is still yours.
Urgent work deserves a clear checkpoint. The mobile concept puts the unresolved question ahead of the approve button.
NORTHBANK / NB-4821
Payment retries
need your review.
Respect cancelled mandates.
The proposed patch stops retrying after a mandate is cancelled. Temporary failures use bounded backoff.
- Checks
- 18 / 18 passed
- Changed files
- 3
- Run cost
- £1.86
Has the cancellation policy been confirmed with the billing owner?
Mobile concept · illustrative review
04 / THE INTERACTION PRINCIPLE
Trust has
a visible scope.
Permission is a product decision. Each level explains what may happen next, and where a person must intervene.
Read the task. Map the constraints.
An agent can inspect the repository and propose a plan. It cannot write or publish.
05 / DECISIONS NEED CONTEXT
One action.
Three questions.
The product makes the instruction, the execution and the unresolved question visible in one place. Select each layer to inspect the sample record.
What was actually asked?
Northbank asked for bounded debit retries after temporary failures. Cancelled mandates must never be retried.
06 / THE VISUAL IDENTITY
A point of view.
A recognizable signal.
Optical texture gives the brand its atmosphere.
Precision gives the interface its structure.
See the whole.
Decide the next.
HUMAN DIRECTION / AGENT EXECUTION
Keep the
human in focus.
VANTAGE / CAMPAIGN STUDY 002THE MARK / CONVERGENCE
Two directions converge into a V. A detached diamond holds the human viewpoint above the work. A single silhouette, legible at app-icon size.
Download SVG mark ↗
THE VOICE / PRECISE, HUMAN
Aa Bb 0123
IBM Plex Sans gives the product and editorial headings a steadier rhythm. Bricolage Grotesque is reserved for the wordmark; Geist Mono labels the operational record.
RUN-4821 / REVIEW / £1.86THE PALETTE / OPTICAL SIGNAL
Carbon and chalk support long reading. Cobalt identifies the system; orange identifies the moment that needs human attention.
#101416 / #F3F0E8 / #246BFD / #F16B3507 / MANY TOOLS, ONE REVIEW CONTEXT
Know where
the work came from.
The source label is part of the evidence. These examples show how familiar tools and model families could be represented in a common record.
Sample connector, model and source contexts only. No live integrations, partnerships or endorsements are claimed.
Anthropic / Claude
A sample reasoning record, with the instruction and supporting documents attached.
- Workspace
- Northbank
- Provenance
- Sample fixture
- Access
- Review only
08 / AN ACCOUNTABLE RECORD
Nothing
disappears
into the run.
When an agent stops, the reason stays attached. When a person decides, the review stays beside the evidence.
FIG. 04 / A HUMAN REMAINS IN VIEWRun paused at policy boundary
Cancelled mandate behavior required an owner decision.
Checks attached to proposal
18 passed; diff and three touched files linked to the run.
Review opened by Maya Chen
Reasoning and checks inspected. Approval remains pending.
09 / WHAT COMES NEXT
Clarity must
survive real work.
The next step is a pilot with people who review agent work every day.
Measure time to a considered decision, evidence opened before approval, and changes reverted after review. Investigate whether reviewers understand the scope of each permission level.
This client project presents a brand system and interactive product prototype. It does not present measured customer outcomes or production integrations.
Build with Aureumgrid ↗