/autonomous-investigation

Autonomous Investigation Protocol — In Action

All companies, products, URLs, and figures are fictional. This example shows the protocol's clauses firing in sequence during one investigation — including the moment most examples skip: the user revising the search plan at the gate, and what an honest gaps list looks like when the world doesn't cooperate. The scenario matches the suite's other examples (Fieldlight, a trades scheduling tool investigating the FSM market).

1. The invocation credits the budget

User: Run an investigation on DispatchCrow's expansion trajectory — this supports our decision on how fast to ship quoting. I already know they raised a Series A in January.

Two of the three budget questions (target, decision) arrived inline; the known fact is credited as context. One question remains in budget, and the skill spends it well:

Agent: One question before I plan the search: is the decision about whether to ship quoting, or when? The evidence bar differs — "whether" needs demand proof, "when" needs their velocity read.

User: When. Whether is decided.

2. The search-plan gate — and a revision

Agent: Search plan (say "revise" to change it):

• Sweep DispatchCrow's changelog, careers page, and blog for shipping velocity and roadmap signals
• Source mix: their site, funding coverage, review sites, practitioner forums
• Facts get URLs; interpretations get labeled Inference; gaps become Assumptions to validate

User: Revise — add their app store release notes. Their web changelog lags the mobile app.

Agent: Good catch — app store release histories added as a SIGINT source. Proceeding.

Ten seconds of plan review just improved the run. That's the gate doing its job — redirecting a plan is cheap; redirecting a finished report is not.

3. Labeled findings, a stack, and an honest gaps list

Key findings:

• Shipped quoting Nov 2025, payments Apr 2026 — five months between module launches — Fact (changelog)
• App store release cadence doubled since the Series A: 2-week cycles now — Fact (release history)
• Two senior dispatch hires from the category incumbent in July — Fact (their blog)
• Payments moved down a pricing tier in July — Fact (pricing archive diff)

The stack (3 channels agree): faster release cadence (SIGINT) + dispatch-specialist hires (HUMINT) + payments repositioned as volume play (SIGINT-pricing) → they are building toward the dispatch-led mid-market now, funded by payments volume — actionable intelligence, not a watch item.

Reading for your decision: their module gap was five months. If capacity planning started at the July hires, expect it Q1 2027 — Inference (their own historical cadence). Your quoting window is roughly two quarters — Inference.

Gaps (couldn't find, not guessed):

• No revenue or payments-volume figures anywhere public — their take-rate economics are unknowable from outside
• No beta signups or docs for capacity planning yet — the Q1 2027 date rests entirely on cadence inference

Final Step — reply 1, 2, 3, 4, a combination, or "Verbose Mode":

1. Set up a competitive-intel-watch tripwire on their capacity-planning beta
2. Deep-dive their payments economics from processor partnerships
3. Trace this timeline into our roadmap assumptions
4. Discovery questions to validate the two-quarter window

4. The same skill, unattended

Three months later the same invocation runs on a schedule. No human answers; the budget is spent silently (context was saved inline), the gate auto-continues with the revised plan (app store sources included — the revision persisted because it was part of the saved invocation), and the output lands in the same schema. The team reads the diff: release cadence held, a "capacity view (beta)" line appeared in the app store notes — the tripwire from option 1 fires, one quarter earlier than the inference predicted.

Why this example works

• Every clause of the contract appears under load. Budget credited by inline context, one question spent surgically, the gate catching a source gap, labels on every claim, a 3-channel stack graded "actionable," a gaps list instead of invented figures, exactly 4 closing options.
• The gaps list is the do-not-invent list working. Payments economics were unknowable, so the report says unknowable — the tempting move (a plausible-sounding take-rate estimate) would have poisoned the roadmap math downstream.
• The unattended run shows why the clauses exist. Question budget, auto-continuing gate, and stable schema are exactly the three properties that let the scheduled run happen at all — and the diff, not the report, is what the team reads.