AI agents · GTM · neurosymbolic patterns

Neurosymbolic GTM Agents —
every AI agent workflow broken down within the Customer Lifecycle

The Neurosymbolic GTM Agent Directory by Bowtie Funnel Lab is an open registry mapping deterministic, multi-agent AI workflows across the entire 7-stage Customer Lifecycle—from Awareness to Expansion. By pairing neural intelligence (LLMs for intent classification and natural language narration) with symbolic execution (100% auditable code, SQL rules, and CRM data pipelines), this architecture eliminates AI hallucinations and optimizes token costs using a tier-right model switchboard. Explore production-ready RevOps, Sales, and Customer Success agent blueprints engineered for zero-error revenue orchestration.
Simple / Medium
Pure sync, classification, scoring, or narration glue. The trigger arrives structured — there is little to frame.
Gemini Flash-Lite · Haiku 4.5 · GPT-5 nano · Gemma 3
Guarded judgment
Structured triggers with a small judgment residue — routing edge cases, fit scoring, hygiene, sourced answers.
Gemini 2.5 Flash · Claude Haiku 4.5 · GPT-5 mini · Qwen 3
Complex · language is the deliverable
The last mile carries the value — long-form drafts, posts, scripts, board commentary within guardrails.
Claude Sonnet 5 · GPT-5 · Gemini 3 Pro · Claude Opus 4.8
Reasoning · the loop carries the framing
Schedule-fired framing and causal work — forecast risk, win/loss attribution, ROAS reallocation.
Claude Sonnet 5 · Gemini 3 Pro · DeepSeek R1 · Claude Opus 4.8
How the switchboard decides: tool-loop reliability & structured judgment → Claude Sonnet 5; massive multi-source context → Gemini 3 Pro; strict-schema drafting → Claude Haiku 4.5; price at classification volume → Gemini Flash / Flash-Lite; deep causal reasoning → DeepSeek R1 / Opus 4.8. Open-weight picks (Gemma 3, Llama 3.3, Qwen 3, Mistral Small 3) fit when data must stay in-house.
Desk 1 · RevOps

32 RevOps agents — what fires them, and how the work really splits

Not every agent is 10-80-10 — the split is a dial. Each section carries its tier's real split: 10-80-10 for reasoning, ~5-90-5 for guarded judgment, ≈5-60-35 when the deliverable is language, 0-100-0 for pipelines. A deterministic trigger is free framing — the more structured the event that wakes an agent, the smaller its neural first mile, the cheaper the model the switchboard picks. Splits are illustrative estimates based on task shape, not measured telemetry.
First mile · Neural

The model reads trigger + intent and frames a structured problem — scope, baseline, drivers. No computing. It only translates the ask into a spec the middle can run.

Middle mile · Deterministic

Where the trustworthy answer is built — rules, arithmetic, queries, API waterfalls, causal attribution against real records. Auditable. Little or no LLM.

Last mile · Neural

The model narrates the computed result in plain business language. It never invented the numbers. Absent when the job is a pure sync — nothing to say.

Swipe to see the full table →

Agent Trigger First mile · frame the problem Middle mile · the real work (deterministic) Last mile · narrate Potential tools Model options (switchboard pick in bold)
Neurosymbolic — full pattern, both neural miles earn their keep10 - 80 - 10
ForecastExplain forecast risk before quarter close 3 days pre-closeSchedule Frame which quarter, segment, baseline, and which drivers to test Pull forecast snapshots + actuals → variance per driver → attribute the gap → check vs. close history Ranked drivers + $ impact + recommended deal review CRM API, forecast snapshots, historical actuals Claude Sonnet 5·Gemini 3 Pro·GPT-5·Claude Opus 4.8·Grok 4·DeepSeek R1
Win/lossAttribute why deals are won and lost Deal marked closed won/lostEvent Frame the deal set + dimensions to attribute (price, competitor, timing, champion) Mine calls + CRM + loss surveys → attribute causes → aggregate patterns → promote recurring causes to rules Why-won / why-lost narrative + pattern callouts Call recordings, CRM close data, loss surveys Gemini 3 Pro·Claude Sonnet 5·GPT-5·Claude Opus 4.8·Grok 4
Churn signalFlag accounts likely to churn, with the cause Weekly usage refresh / usage dropThreshold Frame which accounts, which signals, and the risk window Join usage + tickets + CRM history → risk score with cause → check against past churns Flagged accounts + cause + suggested save play Product usage, support tickets, CRM history Gemini 3 Pro·Claude Sonnet 5·GPT-5·DeepSeek R1
Commit confidenceScore how trustworthy each commit is Rep submits commit / pre-forecast callEvent Frame which commits, scored against what history Compare commit vs. stage history vs. close actuals → trust score → check vs. outcomes Per-commit confidence + why CRM stage history, rep commit log, close actuals Claude Sonnet 5·Gemini 3 Pro·GPT-5·Grok 4·DeepSeek R1
Board reportingCompile board reports on schedule Monthly / quarterlySchedule Frame the period + metric set Pull metrics warehouse → compute → assemble deck deterministically Exec-facing written commentary Metrics warehouse, slides/docs API Claude Sonnet 5·Gemini 3 Pro·GPT-5·Claude Opus 4.8
Guarded judgment — structured triggers shrink the edges~5 - 90 - 5
Deal riskSurface slipping deals and why Deal untouched 14d / stage ageThreshold Minimal — the trigger already names the deal Rules detect slippage deterministically; model narrates why from activity/email/meeting signals; validators gate Short "why it's slipping" note CRM activity, email & meeting signals Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3·Mistral Small 3·Llama 3.3 70B
Expansion signalFlag upsell and expansion opportunities Usage / seat count crosses criteriaThreshold Frame the account + expansion criteria Rules flag against defined usage/billing/seat criteria; model narrates the fit Upsell opportunity note Product usage, billing data, seat counts Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3·Qwen 3
Quote/CPQGenerate quotes inside pricing guardrails Opportunity reaches quote stageEvent Frame requested config + customer Price book + CPQ rules compute the valid price (code); model drafts language within guardrails Quote document Price book, CPQ rules engine, CRM Claude Haiku 4.5·Gemini 2.5 Flash·GPT-5 mini·Mistral Small 3
Lead scoringRank inbound leads by fit New / updated leadEvent Near-zero — payload is structured; map to ICP rubric Enrichment + rubric classification; validators check the score Score + one-line reason Enrichment data, ICP rubric, CRM Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3·Qwen 3·Llama 3.3 70B
CRM hygieneWrite clean fields back to the CRM Record created/edited / nightly sweepEvent Frame which fields + which records Field validation rules; model drafts clean values; validators gate every CRM write Change summary (optional) CRM API, field validation rules Claude Haiku 4.5·Gemini 2.5 Flash·GPT-5 mini·Gemma 3·Mistral Small 3
RenewalSurface renewal risk early Contract date − 90 daysSchedule Frame the account + contract window Rules on dates/usage/tickets trigger; model narrates the risk Renewal risk note Contract dates, usage, support tickets Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
GTM AssistantTriage the inbox, run the calendar, deliver the daily brief Incoming email / daily brief cronEvent Minimal — triage against a learned rubric (VIP senders, ignore rules) decides what matters Calendar availability + scheduling APIs, reminders from email deadlines, brief assembly from calendar + inbox + attendee research; outward replies & accepts gated Daily brief + "what matters" notifications Gmail, Google Calendar, web research API Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3
Account sourcing / TAMBuild & qualify lookalike account lists New segment / ICP refresh · event or keyword list dropSchedule Frame the ICP filters + lookalike seed — or the event/keyword scope Pull from databases + scraping + event attendee / exhibitor lists → dedup into Clay → AI-qualify fit against the ICP rubric Qualified account list + fit reasons Bitscale (waterfall + event/attendee scraping), Clay, Ocean.io, Apollo, Discolike, AI Ark Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
Buying committeeMap decision-makers, champions & influencers per account Account reaches Tier 1/2 · or flagged for expansion / renewalEvent Frame the personas to map per account + the play (net-new, upsell, or renewal) Source contacts (Sales Nav / Apollo / AI Ark) → classify role → dedup into Clay → match a SPOC to the open expansion / renewal play Stakeholder map + persona fit + SPOC per active account Bitscale, Sales Navigator, Apollo, AI Ark, Clay, CRM Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
Account tieringScore accounts into T1/2/3 on an ICP rubric Account enriched / list refresh · Salesforce list watchEvent Near-zero — map the account to the ICP rubric Clay formulas + enrichment score fit → assign Tier 1/2/3 → fill missing firmographics → route ICP-fit accounts to the right rep; rules gate the CRM write Tier + one-line reason + route Bitscale, Clay, HubSpot, Salesforce, enrichment APIs Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3
Signal / intentCapture buying signals, normalize & score intent Signal webhook (1st/2nd/3rd party) · LinkedIn post/comment · competitor G2 reviewEvent Frame the signal type + account Ingest → normalize (domain / LinkedIn URL) → dedup in Clay → score intent against criteria → AI-qualify the engager vs. ICP → auto-push fits into email + LinkedIn sequences Scored signal + suggested play Bitscale (LinkedIn + review-site signals), Clay, RB2B, Warmly, Amplitude, Crossbeam, Instantly/HeyReach Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
Meeting prepBuild a brief before every meeting Meeting T − 1hSchedule Frame attendees + account Retrieval from calendar + CRM + web → assemble the brief One-page brief Calendar, CRM, web search Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
Guarded drafting — the deliverable is language, gated per item≈5 - 60 - 35 · gated
OutboundResearch accounts, draft personalized outbound emails Inbound lead signal from the CRMEvent Near-zero — ICP, personas & voice are standing config from onboarding; the signal is the spec Screen the signal → enrich → research across web / LinkedIn / CRM / warehouse (incl. competitor proof / displacement angle) → dedup vs. past outreach → Gmail-draft rail, per-item gate (outward email never graduates) Personalized email drafts, sources & thought-process surfaced under each (gated) Bitscale, CRM (Salesforce), Instantly, HeyReach, Gmail drafts, Slack, warehouse (BigQuery), enrichment APIs Claude Sonnet 5·GPT-5·Gemini 2.5 Flash·Claude Haiku 4.5
Account battlecardBuild a sales-ready account dossier when a new lead lands New lead in the CRM · @bot account mention in SlackEvent Frame the account + what the rep needs — ICP fit, personas, pain hypotheses, hooks Enrich firmographics → research streams (news, funding, hiring, tech stack, reviews) → map the buying committee → assemble sections → Notion / CRM; review gate Big: the account battlecard — overview, fit, stakeholders, pains, discovery questions & talk track Bitscale / Clay (enrichment), Perplexity / web research, Salesforce, HubSpot, Notion, Slack Claude Sonnet 5·GPT-5·Gemini 3 Pro·Gemini 2.5 Flash
Expansion battlecardRetention, cross-sell & upsell dossier for existing customers Renewal window (T − 90d) · health-score drop · @bot account mention in SlackEvent Frame the account + the expansion motion — renew, upsell seats/tier, or cross-sell a module Pull usage + billing + support + CS health + call notes → score renewal risk & whitespace → 5 parallel Claude generators → merge → Notion / CRM; review gate Big: the expansion battlecard — health, whitespace/upsell, cross-sell fit, save plays, QBR talk track Product analytics (Amplitude/Pendo), billing, Zendesk, Gainsight/Vitally, Gong, CRM, Notion, Slack Claude Sonnet 5·GPT-5·Gemini 3 Pro·Gemini 2.5 Flash
Deterministic pipeline — code computes; a model narrates at most0 - 100 - 0 · 0 - 95 - 5
Call debriefTurn calls into summaries and CRM fields Transcript readyEvent None — transcript arrives in Extraction with schema validation → structured CRM fields (deterministic write) Call summary narration Transcription API, CRM write API Gemini 2.5 Flash-Lite·Claude Haiku 4.5·GPT-5 nano·Gemma 3·Mistral Small 3
Competitive intelAggregate competitor mentions New transcript/news / weeklyEvent None — the competitor roster is standing config Classify mentions at volume across transcripts, news feeds, review sites Digest narration Call transcripts, news feeds, review sites Gemini 2.5 Flash-Lite·Claude Haiku 4.5·Gemma 3·GPT-5 nano·Qwen 3
Lead routingRoute inbound leads instantly New-lead webhookEvent None — the payload is the spec Routing rules assign 95% deterministically (territory/owner); model handles only the residue Route reason on edge cases only Routing rules, territory & owner data, Clay, Slack Gemini 2.5 Flash-Lite·Claude Haiku 4.5·Gemma 3·GPT-5 nano·Qwen 3
EnrichmentFill missing account and contact data Record with missing fields · scheduled stale-contact sweepEvent / Schedule None — record arrives in API waterfall (Clay / Findymail / BetterContact) fills + validates the data → on the scheduled sweep, re-enrich stale contacts, replace bad emails, and flag departed contacts to protect deliverability Mostly none · change / departed summary on sweeps Bitscale (data waterfall), Clay, Findymail, BetterContact, Ocean.io, AI Ark, CRM, HubSpot Gemini 2.5 Flash-Lite·Claude Haiku 4.5·GPT-5 nano·Gemma 3
Website de-anonIdentify anonymous visitors & route them Anonymous visitor identifiedEvent None — visitor record arrives in De-anon vendor (Warmly / RB2B / Albacross) + enrichment → score → route to owner / outbound Slack alert on Tier 1 Warmly, RB2B, Albacross, Clay, Slack Gemini 2.5 Flash-Lite·Claude Haiku 4.5·GPT-5 nano·Gemma 3
Champion Job-Change TrackerRe-route ownership when a champion changes companies Job-change signal on a tracked contactEvent None — the signal arrives structured, naming the person Waterfall finds the new email + company → match to CRM accounts / open deals → rules flag re-route + suggest a re-engagement play; validators gate the write Slack alert on re-route + owner handoff Bitscale waterfall, job-change signals, CRM (Salesforce/HubSpot), Slack Gemini 2.5 Flash-Lite·Claude Haiku 4.5·Gemma 3·GPT-5 nano
AttributionTie outbound touches back to pipeline & revenue Outbound touch / signupEvent None — the touch arrives structured OutboundSync logs every touch → HubSpot property writes + deal-stage rollups (deterministic) Small model explains the report OutboundSync, HubSpot code only·Haiku 4.5 / Gemma 3 to narrate
Multi-threadFlag single-threaded deals Deal review / weeklySchedule None SQL contact count per deal → flag single-threaded Small model narrates the flag CRM contacts query (SQL) code only·Gemma 3 / Haiku 4.5 / Flash-Lite to narrate
Pipeline coverageFlag coverage gaps against quota WeeklySchedule None Arithmetic — pipeline $ vs. quota Narrate the gap CRM pipeline data, quota table code only·Gemma 3 / GPT-5 nano to narrate
Coverage & capacityTrack capacity and coverage vs. target WeeklySchedule None Arithmetic — rep roster / quota vs. target Narrate Rep roster, quota table code only·Flash-Lite / Haiku 4.5 to narrate
Comp & quotaModel quotas and attainment Comp period / deal closed-wonEvent None Deterministic money math — comp plan rules + closed-won data Small model explains the statement Comp plan rules, closed-won data code only·Haiku 4.5 / Gemma 3 to explain
Activity captureAuto-sync email and meetings to the CRM Email sent / meeting occursEvent None Pure integration sync — email / calendar → CRM None — no model Email / calendar APIs, CRM API code only — no model
Reading a row: the trigger wakes the agent → the first mile turns trigger + intent into a structured problem → the middle mile computes the trustworthy answer in code/rules → the last mile narrates it. The section badge shows each tier's real split — the neural miles shrink as jobs get more structured, and the switchboard drops the model tier with them. Where a neural mile reads "none", the trigger arrived already structured or the job is a pure sync with nothing to say — that's the discipline, not a gap.
Desk 2 · Marketing

18 marketing agents — what fires them, and how the work really splits

The split is a dial, and marketing turns it differently: when the deliverable is language, the last mile grows — the draft is the product, and the switchboard routes it up a tier. Only paid ads closes its own loop and earns the neurosymbolic tier; for every other agent a human closes the loop, and every external action — post, publish, PR, send, spend — stops at a human gate. Splits are illustrative estimates based on task shape, not measured telemetry.
First mile · Neural

The model reads trigger + intent and frames the work — topic, angle, audience, query set. Standing briefs and structured triggers shrink this toward zero.

Middle mile · Deterministic

Context retrieval, keyword & thread discovery, relevance filters, scoring, generation APIs, publish rails, budgets — and the HITL gate that clears every external action.

Last mile · Neural

In marketing this mile is often the deliverable itself — the post, article, script, or reply. Bigger than in RevOps, but still bounded by guardrails and the gate.

Swipe to see the full table →

Agent Trigger First mile · frame the work Middle mile · deterministic spine Last mile · the deliverable Potential tools Model options (switchboard pick in bold)
Neurosymbolic — the one loop marketing closes itself10 - 80 - 10
Paid adsReallocate spend across campaigns against ROAS Daily pacing / ROAS checkSchedule Frame which campaigns, budgets, and the ROAS target window Pull spend + conversions + pipeline → compute ROAS per campaign → reallocate against rules → check vs. outcomes Reallocation decisions + rationale — budget rails gate the spend Ad platform APIs (LinkedIn/Meta/Google), attribution, CRM pipeline Claude Sonnet 5·Gemini 3 Pro·GPT-5·DeepSeek R1
Guarded analysis — audits & scores are code, recs are judgment≈5 - 85 - 10 · gated
SEOFind, prioritize, and recommend ranking fixes Daily / weekly audit runSchedule Near-zero — page set, keywords, and baseline (last crawl) are standing config; the schedule is the spec Crawl audit + Core Web Vitals + keyword gap + backlinks → prioritize fixes by impact rules → gate before any fix ships Severity-tagged fix recommendations with rationale Site crawler, Lighthouse/CWV, SERP & keyword APIs, backlinks API, CMS / GitHub Claude Sonnet 5·Gemini 3 Pro·GPT-5·DeepSeek R1
GEOTrack & grow brand visibility in AI answers Daily visibility runSchedule Near-zero — the buyer-question query set is standing config, refreshed only when positioning shifts Fan-out probes across engines → citation / mention scoring → deltas vs. last run Ranked recommendations to win citations Multi-model probes, AI-Overview checks, citation scoring Gemini 3 Pro·Claude Sonnet 5·GPT-5
Competitor intelMaintain the battlecard, answer competitor questions in Slack Slack @mention / weekly refreshOn-demand Near-zero — the competitor roster is standing config; the question or the schedule is the spec Weekly: web research per competitor → refresh wiki + Notion battlecard. On ask (any company): retrieve battlecard + fresh search → cite sources Sourced answers in-thread + concise research reports Tavily search, Slack, Notion battlecard Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Qwen 3
Guarded drafting — the deliverable is language, the last mile grows≈5 - 45 - 50 · gated
Battlecard builderGenerate 9-section competitive battlecards + cited Slack Q&A (Klue / Crayon alternative) @bot competitor mention in SlackEvent Detect the competitor, fuzzy-match the registry, route — build a new card or answer from the existing one 5 parallel research streams (Perplexity + Jina scrape) → 5 parallel Claude generators (overview, positioning, features, SWOT, attack surfaces) → merge → Notion + Data Table; review gate Big: the 9-section battlecard + cited Slack answers Slack, Perplexity, Jina AI, Notion, n8n Data Table Claude Sonnet 5·GPT-5·Gemini 3 Pro·Claude Haiku 4.5
WriterDraft long-form articles inside brand & SEO guardrails Weekly cron / keyword handoff from SEOSchedule Pick topic + outline from keyword data Context retrieval (product, SERP), SEO lint & schema checks, OG-image gen, CMS publish rails, review gate Big: the 800–2,000-word draft — the deliverable is the language Keyword input, OG-image model, CMS publish API Claude Sonnet 5·GPT-5·Gemini 3 Pro·Claude Opus 4.8
LinkedInDraft founder-voice posts daily Daily cronSchedule Near-zero — a standing brief + brand voice carries the framing Voice context assembly, dedup vs. past posts, scheduling, OAuth publish rail, review gate Big: the post itself Brand-voice context, LinkedIn OAuth Claude Sonnet 5·GPT-5·Gemini 2.5 Flash·Claude Haiku 4.5
X (Twitter)Scout the timeline, draft posts & threads in your voice Daily cronSchedule Near-zero — standing brief Scout scan (timeline, mentions, web) → voice-profile retrieval, dedup, scheduling, OAuth publish rail, review gate Big: 5 draft posts & threads, humanizer pass Voice-profile KB, timeline/web scout, X OAuth Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3
Hacker NewsPitch launches in builder voice (3 variations) You ask, at a launch momentOn-demand Frame the launch angle from your ask Product context retrieval, 3-variation structure, no auto-post by design Big: three pitch variations Product context; manual posting only Claude Sonnet 5·GPT-5·Gemini 2.5 Flash
RedditFind relevant threads, draft non-promo replies Daily scanSchedule None — keyword config is the spec API discovery → relevance filter drops weak fits → thread metadata (subreddit, age, upvotes) Why-relevant note + drafted reply (gated, posted manually) Reddit API / search, keyword config Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini
Social listeningMonitor brand & keyword mentions, draft outreach Brand / keyword mention detectedEvent Near-zero — keyword config is the spec Monitor mentions (Jungler / Teamfluence) → enrich + score engaged accounts in Clay → dedup Why-relevant note + drafted outreach (gated) Jungler, Teamfluence, Clay, Slack Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini·Gemma 3
CodingTurn technical SEO fixes into GitHub PRs Handoff from the SEO agentEvent None — the finding arrives structured Read repo → generate fix → branch; opening the PR is the gate PR description + diff rationale GitHub API, repo access Claude Sonnet 5·GPT-5·Gemini 3 Pro
Link buildingFind prospects, draft outreach Weekly prospect runSchedule Frame target pages + anchor priorities Prospect data pull → fit scoring by rules → dedup vs. past outreach Personalized outreach drafts (gated) Prospect / backlink data, email outreach Gemini 2.5 Flash·Claude Haiku 4.5·GPT-5 mini
InfluencerRun creator campaigns brief → payout You brief a campaignOn-demand Turn your brief into a structured campaign spec Creator DB matching, rate benchmarks, deliverable tracking, payouts — all code Outreach + negotiation drafts (gated) Creator database, outreach channel, payment rail Claude Haiku 4.5·Gemini 2.5 Flash·GPT-5 mini
YouTubeScript, produce & publish channel videos Weekly cronSchedule Frame the topic from performance data Production pipeline — video gen / edit APIs, thumbnail, upload via Data API, review gate Big: script + title + description Video gen / edit APIs, YouTube Data API Claude Sonnet 5·GPT-5·Gemini 3 Pro
Deterministic pipeline — code or a generation API does the work0 - 100 - 0 · 0 - 95 - 5
UGC videoTurn briefs into short video ads You request a videoOn-demand Brief → render prompt (pure translation) Text-to-video API (~10 min render), storage, credit accounting None — the MP4 is the deliverable Text-to-video API (Veo / Kling / Runway), storage Claude Haiku 4.5·Gemini 2.5 Flash·GPT-5 mini·+ the video model does the real work
RepurposingAtomize one asset into many formats & channels New asset publishedEvent None — the source asset is the spec Templated atomization into per-format variants, scheduling rails, review gate Per-format rewrites (glue) Content archive, format templates, channel APIs Claude Haiku 4.5·Gemini 2.5 Flash·GPT-5 mini
NewsletterAssemble & send the newsletter on schedule Weekly send dateSchedule None — send date + content archive is the spec Assemble from published content, list management, ESP send rails, review gate Subject line + intro glue only ESP API, content archive, subscriber list Claude Haiku 4.5·Gemini 2.5 Flash-Lite·GPT-5 nano
Reading a row: the trigger wakes the agent → the first mile frames the work (standing briefs and structured triggers shrink it to nothing) → the middle mile retrieves, filters, scores, and holds the publish rails → the last mile is the draft. Rows marked "Big" are agents whose deliverable is the language itself — there the last mile carries the value and the switchboard routes up to a stronger writing model. The one neurosymbolic exception is paid ads — it owns a measurable ROAS loop end-to-end, though its budget rails still gate the spend. The constant across both desks: every external action (post, publish, PR, send, spend) stops at a human gate.
Built by Bowtie Funnel

Our own tools

Reusable code we ship in the open — drop-in building blocks for GTM agents, separate from the vendor stack above.

llm-switchboard

local · zero-dep · MIT

A ~130-line prompt router that sits before OpenRouter. It classifies a prompt in <1ms, picks the model, and hands the ID to your OpenRouter call — cheap/free models for the workflow miles, frontier for thinking. The tier picks mirror our RevOps & Marketing agent maps.

prompt→switchboard · SIMPLE · MEDIUM · COMPLEX · REASONING→OpenRouter→model
TierBold pickFree option
SIMPLEgoogle/gemini-2.5-flash-litegoogle/gemma-4-31b-it:free
MEDIUMgoogle/gemini-2.5-flashopenai/gpt-oss-20b:free
COMPLEXanthropic/claude-sonnet-5nvidia/nemotron-3-ultra-550b:free
REASONINGanthropic/claude-sonnet-5nvidia/nemotron-3-…-reasoning:free

Option lists also carry GPT-5 / Gemini 3 Pro / Opus 4.8, plus Kimi K3 in the agentic tiers. route(prompt, {preferFree:true}) swaps to the free model.

View on GitHub → btf-switchboard.js
One switchboard, both desks — the job shape picks the model, and nothing ships ungated.