Sales Outreach Agent

Sonnet 4's tone control is measurably better than GPT-4o on short-form outbound — less AI-voice, better signal integration. Matters because reply rates correlate directly with perceived authenticity.

Alternatives: GPT-4o, Gemini 2.5 Pro

Haiku 4 ranks signals per lead in <600ms for a fraction of a cent. Running Sonnet here is wasteful.

Alternatives: GPT-4o-mini, DeepSeek R1

Firecrawl handles JS-rendered sites and clean markdown extraction. Exa gives semantic search for 'companies that just did X' queries. Combine them.

Alternatives: Serper, Tavily, Custom Playwright

Sender warmup and rotation are not a side project. Dedicated ESPs for outbound cost $100–500/mo and save you from domain blacklists.

Alternatives: Lemlist, Apollo, Custom + SendGrid

Attribution, deal tracking, and AE handoff live in CRM. The agent writes back activities + reply classification as custom properties.

Alternatives: Attio, Folk, Airtable

Per-lead pipeline with retries, rate limits, and per-domain send caps fits event-driven orchestration naturally. LangGraph overcomplicates a fundamentally linear flow.

Alternatives: Temporal, Plain cron + queue

You must A/B test subject lines and openers continuously. Reply rate is the only metric that matters and it varies by ICP.

Alternatives: Humanloop, Langfuse

Volume vs reply rate

Sending 10k messages/day with shallow research drops reply rates below 3% and burns domains. Sending 500/day with real 60s-of-research per lead gets 15–30% reply rates and preserves sender reputation. Modern outbound is depth over breadth — the agent exists to enable depth at scale, not to spray further.

Grounded vs creative writing

A strict 'cite a source' rule makes messages credible but can make them boilerplate if the signal is weak. Better: allow creative openers only when signal confidence > 0.8, otherwise fall back to a clean, short problem-statement template. Worst outcome is fabricated specifics.

Auto-reply vs human handoff

Letting the agent handle 2–3 reply turns beats immediate human handoff for nurture replies. For any reply mentioning pricing, meetings, or signing, hand off immediately — over-committing on price or timeline is the #1 way an agent loses a deal.

Message references a person who left the company

Mitigation: Require every named-person reference to include a source URL and a 'last_seen' date within 30 days. If LinkedIn data is older than 30 days, refetch before composing. Add a final guard: cross-check named title against current company page.

Deliverability collapse from over-sending

Mitigation: Hard cap 50 sends/day per inbox, 200/day per domain, pause any inbox with bounce rate >3% or spam complaint >0.1%. Run continuous warmup on every inbox. Rotate across 5–10 domains for volume.

Agent hallucinates a joint connection or event

Mitigation: Whitelist only these claim types: job change, funding, product launch, hire, content they published. Block all 'we met at', 'our mutual friend', and 'your competitor X' — these are where hallucinations land hardest.

Reply handler commits to pricing or contract terms

Mitigation: Maintain an explicit stoplist of topics (pricing, contracts, legal, timelines beyond 'this week'). Any reply matching these triggers immediate human handoff with the entire thread. Never allow the agent to propose a number.

GDPR / CAN-SPAM violations

Mitigation: Honor unsubscribe within 10 minutes via ESP API. Never email EU leads without documented lawful basis. Maintain a suppression list synced across all inboxes. Include physical address + one-click unsubscribe in every email — not optional.