AI models & usage

This page answers one question: where is AI actually used across PulseAd, and where is it deliberately not? It maps every model to the agents that run on it, flips that to every agent and the model behind it, and is filterable by provider or text. The data is code-verified.

At a glance

PulseAd doesn’t run on one model — it runs on a deliberate split by service:

• ai-api (chat, reports, planning) → OpenAI GPT‑5 family (gpt-5, gpt-5.5, gpt-5.2, gpt-5-mini, gpt-5-nano, gpt-4.1/-mini/-nano), with one Google model (gemini-2.5-flash-lite) for search-term segmentation.
• pulsead-agents (autonomous campaign ops) → Anthropic Claude on AWS Bedrock — Haiku 4.5 by default, Sonnet 4.6 for the decision/writer agents, and Opus 4.8 for the one job judged worth it: writing the weekly report’s summaries.
• AOP (the operator CLI + its fleet runtime) → a third posture, decided independently of both: claude-opus-4-7 pinned on four operator profiles, and OpenAI gpt-5.6-terra via OpenRouter as the fleet default the rest inherit.
• Snowflake Cortex (the semantic data layer) → Claude Sonnet 4.6.
• Creative → Google Gemini for images by default, with OpenAI gpt-5.5 + image_generation selectable per run in both image apps, plus a multi-vendor video gateway (Sora 2, Runway, Kling, Google Veo, Grok).

The pattern: cheap, fast models for classification and collection; stronger models for decisions and writing. Router and monitors run on the lightest models; budget decisions, report writing, and SQL generation run on the strongest.

Explore it

Toggle between by model → agents and by agent → models, filter by provider or surface, or search. Each agent carries a role badge: read (analyzes only), propose (suggests a change for approval), write (executes, gated), internal (writes only to PulseAd’s own state), and generate (produces creative).

Three surfaces, counted separately and never summed. Pulson is the customer product. AOP is the operator CLI staff run — its own agents, its own model choices, no customer request reaches it. Platform is the jobs that build the data Pulson reasons over. Filter to one to see it alone.

A few AOP profiles pin no model of their own; they inherit the fleet default, and their model chip is shown dashed and italic to say so.

40Pulson agents

10AOP agents

2Platform agents

26models

7providers

Where the AI runs

Anthropic 25 OpenAI 24 Google 5 xAI 1 Runway 1 Kling 1 Exa.ai 1

By model By agent

All PulsonAOPPlatform

All OpenAIAnthropicGooglexAIRunwayKlingExa.ai

gpt-4.1 OpenAI 1 agents

• Report streaming / metadata Pulson Reports read Report narrative text and metadata

gpt-4.1-mini OpenAI 2 agents

• Action Router Pulson Chat read Classifies each message and resolves context
• Image / PDF summarizer Pulson Utility read Vision extraction of uploaded attachments

gpt-4.1-nano OpenAI 1 agents

• Data summarizer Pulson Utility read Misc. structured data summaries

gpt-5 OpenAI 1 agents

• Action sub-agent Pulson Chat write The only agent that changes an Amazon account (gated)

gpt-5-mini OpenAI 2 agents

• Task sub-agent Pulson Chat internal Manages the internal inbox & tasks
• Brand augment extractor Pulson Chat internal Pulls missing brand fields out of the user's message

gpt-5-nano OpenAI 2 agents

• Memory classifier Pulson Chat internal Background — builds durable brand memory (16-card)
• Chart builder Pulson Chat read Turns a result set into a chart spec

gpt-5.2 OpenAI 2 agents

• PSM researcher / Pulson engine Pulson Chat read Brand-facing chat engine (Growth Twin)
• Dashboard Report Agent Pulson Reports read The report engine — fetch data, plan directives

gpt-5.5 OpenAI 5 agents

• Query sub-agent Pulson Chat read Answers about your own performance from Snowflake
• Insight sub-agent Pulson Chat read External market / competitor / keyword lookups
• Action sub-agent Pulson Chat write The only agent that changes an Amazon account (gated)
• Composer Pulson Chat read Writes the final user-facing reply
• Performance Diagnose Pulson Reports read Finds declining campaigns, recommends actions

gpt-5.5 (image_generation) OpenAI 2 agents

• banana-bunch (image gen) Pulson Creative generate Product image + prompt → many images; model is selectable
• ads-variation (banner) Pulson Creative generate PSD → sized banner variations; model is selectable

gpt-5.6-terra OpenAI 6 agents

• bid-manager AOP Ads operations propose Bid adjustments across campaigns
• report-writer AOP Ads operations read Writes operator-facing reports
• orchestrator AOP General internal Routes work to the other profiles
• librarian AOP General read Serves the Pulse Library docs to the operator
• growth-supporter AOP General read General growth questions
• psm-onboarding-facilitator AOP General read Walks a new PSM through onboarding

Sora 2 OpenAI 1 agents

• video-generator (gateway) Pulson Creative generate One brief routed across video providers

claude-3-5-sonnet-20241022-v2 Anthropic 1 agents

• SA tutorial extractor Platform Semantic layer build internal Regenerates semantic assets from tutorials

claude-haiku-4-5 Anthropic 10 agents

• monitoring (Campaign Health) Pulson Autonomous ops read Daily anomaly detection
• diagnoser (Campaign Health) Pulson Autonomous ops propose Root-cause; proposes structural actions
• rtb_analyst (Campaign Health) Pulson Autonomous ops propose RTB rule-group tuning proposals
• executor (Campaign Health) Pulson Autonomous ops write Applies approved bid/budget/pause actions
• oc_monitor (Optimize Cycle) Pulson Autonomous ops read Collects 7-day + MTD performance
• ads_scheduler Pulson Autonomous ops read Parses NL request into a scheduled task
• ads_executor Pulson Autonomous ops write Runs due scheduled changes (gated)
• amc_router Pulson Autonomous ops read Routes AMC requests to advisor/writer
• amc_advisor Pulson Autonomous ops read Answers conceptual AMC questions
• Column description generator Platform Semantic layer build internal Writes column descriptions for the data model

claude-opus-4-7 Anthropic 4 agents

• campaign-planner AOP Ads operations propose Plans campaign structure and launches
• pacing-analyst AOP Ads operations read Watches spend pacing against target
• targeting-manager AOP Ads operations propose Keyword and target curation
• tool-maintainer AOP Ads operations internal Maintains the operator's own tool catalog

claude-opus-4-8 Anthropic 1 agents

• weekly-report summary Pulson Autonomous ops read Writes each section's summary (replaced the 4-agent pipeline)

claude-sonnet-4-6 Anthropic 8 agents

• oc_input_checker (Optimize Cycle) Pulson Autonomous ops read Validates brand prerequisites
• oc_rebalancer (Optimize Cycle) Pulson Autonomous ops propose Computes & proposes product-group budgets
• oc_optimizer (Optimize Cycle) Pulson Autonomous ops propose Proposes campaign-level allocation
• oc_executor (Optimize Cycle) Pulson Autonomous ops write Applies budgets to Amazon (KISS-only live)
• oc_verifier (Optimize Cycle) Pulson Autonomous ops write Verifies & auto-corrects budget drift
• amc_writer Pulson Autonomous ops read Generates & saves AMC SQL (no execute)
• keyword_harvester Pulson Autonomous ops propose Proposes profitable exact-match keywords
• waste_keyword_pauser Pulson Autonomous ops propose Proposes pausing zero-click spend

claude-sonnet-4-6 (Cortex) Anthropic 1 agents

• AMAZON_ADS_AGENT (6 analysts) Pulson Semantic SQL read Text-to-SQL analytics over the data model

gemini-2.5-flash Google 1 agents

• banana-bunch (prompt extract) Pulson Creative read Reverse-engineers a prompt from one uploaded image

gemini-2.5-flash-lite Google 1 agents

• Search-term segmentation Pulson Utility read Classifies search terms + brand recognition

gemini-3-pro-image-preview Google 1 agents

• banana-bunch (image gen) Pulson Creative generate Product image + prompt → many images; model is selectable

gemini-3.1-flash-image-preview Google 1 agents

• ads-variation (banner) Pulson Creative generate PSD → sized banner variations; model is selectable

Google Veo Google 1 agents

• video-generator (gateway) Pulson Creative generate One brief routed across video providers

Grok xAI 1 agents

• video-generator (gateway) Pulson Creative generate One brief routed across video providers

Runway Runway 1 agents

• video-generator (gateway) Pulson Creative generate One brief routed across video providers

Kling Kling 1 agents

• video-generator (gateway) Pulson Creative generate One brief routed across video providers

Exa.ai (search) Exa.ai 1 agents

• News & Market Intelligence Pulson Planning read News briefings (Exa.ai search + LLM summary)

Where AI is not used (by design)

Just as important for a product leader: the execution and safety paths are deterministic, not AI. That’s intentional — it makes behavior predictable and auditable. No LLM decides these:

• Scope & auth — resolving team + country, role checks, the 404-hides-existence rule.
• The approval & apply machinery — the action-request lifecycle, the approval gates, the dispatch to Amazon, the viewer-role block. AI proposes; deterministic code applies.
• Budget math — the rebalance calculator is a pure function, not a model.
• The KISS-only live whitelist and per-brand amazon_ads_mode — hard-coded gates.
• Report serialization — the report’s final render step emits fixed directives with no LLM (the LLM only plans the content).
• Memory write-back — the classifier is AI, but the dispatcher/effects that persist memory are deterministic.
• Data plumbing — Snowflake’s SELECT/WITH-only guard, off-Amazon ingestion (Apps Script + SQL), the report aggregator endpoints.
• Retrieval ≠ generation — Exa.ai and Tavily fetch sources; the LLM only summarizes them.

How to read this

• AI is concentrated in reasoning and content: understanding a request, analyzing performance, writing reports, proposing changes, generating creative.
• AI is kept out of execution: anything that actually changes an Amazon account, moves money, or enforces a rule runs on deterministic code behind a human gate.
• The model choice signals intent: a Haiku/gpt-*-mini/-nano agent is a cheap classifier or collector; a Sonnet/gpt-5/gpt-5.2 agent is a decision or writing step.

How we got here

This page is today’s snapshot. For the path — when each service came online, every model swap since, and what was retired — see Model history. Two things it makes visible that a snapshot cannot: the vendor split has held since each repo’s first commit and no agent has ever crossed it, and upgrades consistently hit the reasoning seats while the cheap classifier seats stay put for months.

Models and assignments change as the product evolves — re-run /sync-docs to refresh this from source.