Research · Case study · 6 September 2026
WhoFollows: scoped agent access to verifiable results
Can a small, read-only interface help an agent explain changes accurately while keeping the application in charge of the facts?
WhoFollows compares user-provided follower and following lists. For the WebMCP Challenge, we added an experimental agent interface to the existing product. Submission was confirmed on 3 September; organisers announced that submissions had closed on 4 September. Participation does not imply an award or endorsement.
The problem
An unchanged following count can hide a changed list. WhoFollows saves snapshots and computes their differences using deterministic rules. A change in a list does not explain anyone’s motives.
Our hypothesis
Giving an agent a structured comparison result through a narrowly defined tool may improve factual accuracy and reduce unsupported explanations. The application computes the changes; the agent communicates the result. WebMCP is an optional read-only interface, with no embedded language model or autonomous Instagram access.
What the demo demonstrates
The demo uses two synthetic Following snapshots for @demo.profile. The documented successful check on 3 September recorded a call to get_latest_changes and the following result:
| Measure | Result |
|---|---|
| Added | @new.account |
| Removed | @older.account |
| Unchanged | 3 accounts |
| Following count | 4 → 4 |
This demonstrates the feasibility of one specific path: an agent calls the application’s tool and receives its comparison result. An earlier run that displayed the right changes without calling WebMCP is excluded as evidence of integration.
Evidence scope: this account relies on the retained verification history, not a fresh independent replication. An invocation recording and an exact demo revision are needed for an independently reproducible evidence package.
What we do not claim
The demo does not establish universal browser or agent compatibility, repeated-run reliability, whole-product security, or an advantage over the ordinary interface. It does not validate the completeness of real user lists. Reduced anxiety and improved usability remain untested hypotheses.
We do not claim scraping, autonomous access to Instagram, account control, or knowledge of why someone followed or unfollowed. The example contains synthetic data. A read-only tool alone is not proof of complete system privacy.
The next experiment
First, preserve the challenge version, synthetic inputs, expected outputs and an actual invocation recording. Then run a small paired comparison using the same model and environment:
- Page access: the agent reads the comparison displayed on the page, without WebMCP.
- Tool access: the agent retrieves the same result through
get_latest_changes.
Proposed pilot: 10 predefined synthetic cases, 3 independent runs per case, in each mode — 60 runs in total. Include no change, additions, removals, replacement at the same count, and a missing previous snapshot. Missing history must mean “comparison unavailable”, not “nothing changed”.
Measure factual accuracy, unsupported assertions and task completion. Record actual tool calls separately, reset context between runs, alternate mode order and retain failures. Equal performance would also be informative. This pilot would assess the interface in one environment; understanding and emotional effects require a separate study with people.
Status: technical demonstration documented; comparative experiment proposed, not yet conducted.