Launch Filter

Agent Launch Filter

Takes any agent announcement — URL, press release, release notes, or pasted text — and runs it through a sequenced 5-question filter to produce a scored verdict and a go/wait/skip recommendation.

Replaces a manual weekly evaluation loop with a 30-second invocation.

The 5-Question Filter

Questions are sequenced: each answer constrains the next. A hard "no" on #1 or #3 is a skip signal regardless of later scores.

Phases

Phase 1: Gather the Announcement

If the user provided a URL, fetch the page content (WebFetch). If they pasted text, use it directly. If neither, ask: "Paste the announcement text or give me a URL."

Extract:

• Product/agent name and vendor
• Key capability claims
• Integration points mentioned
• Data access claims
• Pricing model (if stated)

Phase 2: Run the Filter

Score each question 0–2:

• 2 = clear yes with evidence
• 1 = partial / ambiguous / "coming soon"
• 0 = no / opposite is true

Q1 — Tool Integration

• 2: drops into existing stack (MCP, API, OAuth, no data migration required)
• 1: partial integration, some migration needed or new credentials/onboarding required
• 0: requires platform switch, data export, or proprietary lock-in to access features

Q2 — Openness

• 2: open SDK or MCP exposure; other agents can call it programmatically
• 1: API exists but limited (rate limits, no SDK, closed beta)
• 0: closed; no programmatic access for external agents

Q3 — Data Access ← hard gate: score 0 = skip unless Q1+Q2 compensate

• 2: natively accesses data graphs the team uses (M365, Google Workspace, Salesforce, existing DBs)
• 1: accesses some relevant data but requires connectors or sync jobs
• 0: no access to team data; operates only on inputs you manually provide

Q4 — Ecosystem

• 2: marketplace listed, SDK available, named partners announced, ships regularly
• 1: early marketplace or SDK, few partners, infrequent releases
• 0: no marketplace, no SDK, no visible partner traction

Q5 — Stackability

• 2: agent A can invoke agent B via API, MCP, or SDK; deterministic hooks available
• 1: some scripting possible but not first-class; workarounds needed
• 0: silo; no path for other agents to build on top

Phase 3: Score and Recommend

Total score: 0–10

Routing buckets:

Override rules:

• Q3 = 0 AND (Q1 + Q2) < 3 → hard Skip regardless of total
• Q1 = 0 → bump recommendation one step toward Skip (Go → Wait, Wait → Skip)

Phase 4: Output

## Launch Filter: [Product Name]

**Verdict: [Go / Wait / Skip]**   Score: [N]/10

| Question | Score | Evidence |
|---|---|---|
| Q1 Tool Integration | [0-2] | [one sentence] |
| Q2 Openness | [0-2] | [one sentence] |
| Q3 Data Access | [0-2] | [one sentence] |
| Q4 Ecosystem | [0-2] | [one sentence] |
| Q5 Stackability | [0-2] | [one sentence] |

**Rationale:** [2-3 sentences on why this verdict, what the key signal was]

**If Go:** [Concrete next step — what to integrate, where, how long]
**If Wait:** [What to watch for — specific signals that would move to Go]
**If Skip:** [Alternatives worth evaluating instead, if any]

Verification

• Every question has a score 0–2 with at least one evidence sentence
• Override rules applied if Q3 = 0 or Q1 = 0
• Recommendation matches the score bucket (or override is documented)
• Output is copyable — no raw tool output in the block

Source

Nate Kadlac, "The 5-question filter I run every agent launch through (so you can stop reading release notes)", natesnewsletter.substack.com, 2026-04-29.