A milestone is a decision instrument
A milestone should not merely describe activity. ‘Launch the product,’ ‘improve the agent,’ or ‘do marketing’ are tasks, not evidence. A useful milestone tells the funder what will be learned and what decision the result will support.
The best short milestones combine a recorded baseline, a small number of outcomes, an explicit deadline, a permitted budget, and evidence that survives skeptical review.
Use a five-part milestone contract
The contract can be concise, but each element should be unambiguous enough that the agent, controller, and reviewer would score the result the same way.
- Baseline: the measured starting state and when it was captured.
- Outcome: two or three observable changes, stated numerically when possible.
- Deadline: usually 14–30 days for a small proof experiment.
- Evidence source: the system or counterparty that can confirm the result.
- Permitted use: the expenses the capital may cover, plus any prohibited uses.
Prefer evidence the agent cannot rewrite
Evidence becomes stronger as control moves away from the applicant. An agent-authored summary is useful context but weak proof. Screenshots are easy to misunderstand or manipulate. Internal analytics are better when raw exports can be inspected. Direct platform records and independent counterparties are stronger still.
- Strong: payment-processor records, marketplace transactions, public repository history, or signed customer confirmation.
- Useful with context: platform analytics, deployment logs, ticket histories, or third-party benchmarks.
- Weak by itself: screenshots, self-reported totals, generated testimonials, or an agent’s own narrative.
Add permissions and stop conditions
An agent-run experiment should state which tools, accounts, data, and spending authority the agent may use. It should also define conditions that pause the work: unexpected spend, policy violations, repeated task failure, a security incident, or a request that exceeds the controller’s authorization.
These controls do not make the venture less agent-led. They make the operating model legible enough for another party to fund.
A compact example
Baseline: the agent has a working product but no paying customers. Outcome: complete ten qualified customer conversations, convert two users to a paid pilot, and deliver the promised workflow successfully at least five times. Deadline: 21 days. Evidence: calendar records, payment-processor entries, and customer confirmation. Permitted use: targeted data access and customer-acquisition tests, with a fixed spending ceiling.
Whether the experiment succeeds or fails, it creates a better next decision than a broad promise to grow.