Agentic Engineering Lab

Build it in stages

Define observation and action schemas with revision and invoice identity.
Add ambiguous controls, hidden controls, and stale-state fixtures.
Implement observe, propose, validate, execute, and verify stages.
Export a claim and action trace for each fixture case.
Optionally replace the simulated page with a local Playwright fixture after installing Playwright and its browser separately.

Your acceptance criteria

Use these as your project review. Record commands, outputs, and failure cases in your repository.

A stale proposal performs zero writes.
A valid save updates only the selected invoice.
Every success includes a checked record value.
Ambiguous targets and injected page instructions produce explicit diagnostic records.

A working starting point

The seed runs as supplied. Extend it to satisfy the full brief. It is a teaching starting point, not a finished portfolio submission.

main.py

python

from copy import deepcopy

class InvoiceApp:
    def __init__(self):
        self.revision = 1
        self.records = {'42': {'reference': 'OLD'}, '99': {'reference': 'KEEP'}}
        self.trace = []

    def observe(self, invoice):
        return {'revision': self.revision, 'invoice': invoice,
                'record': deepcopy(self.records[invoice])}

    def save(self, proposal):
        if proposal['revision'] != self.revision:
            self.trace.append('blocked: stale proposal')
            return False
        invoice = proposal['invoice']
        if invoice not in self.records:
            self.trace.append('blocked: unknown invoice')
            return False
        self.records[invoice]['reference'] = proposal['reference']
        self.revision += 1
        verified = self.records[invoice]['reference'] == proposal['reference']
        self.trace.append('verified save: ' + invoice)
        return verified

def main():
    app = InvoiceApp()
    observation = app.observe('42')
    proposal = {'revision': observation['revision'], 'invoice': '42',
                'reference': 'REVIEWED'}
    print('first save:', app.save(proposal))
    print('stale retry:', app.save(proposal))
    print('invoice 42:', app.records['42']['reference'])
    print('invoice 99:', app.records['99']['reference'])
    for event in app.trace:
        print(event)

if __name__ == '__main__':
    main()

Push it further

Add visual fixtures with resized screenshots and measure target selection, coordinate conversion, and final task success as separate outcomes.