IMA Framework & IMA-Hermes-Agent
An autonomous 5-agent fleet built on Hermes Agent that automates YouTube creator discovery, transcript hook extraction, and hyper-personalized cold outreach at scale.
Problem & Context
Before IMA, creator discovery and cold outreach in influencer marketing were manual, fragmented, and notoriously unscalable. Agencies relied on human account managers spending 15 to 20 minutes per creator manually searching YouTube, watching video intros, skimming transcripts to find relevant content hooks, writing bespoke cold emails, and logging entries into static spreadsheets.
This workflow presented three major bottlenecks:
- Low Volume: A single operator could thoroughly analyze and outreach to at most 15–20 creators per day.
- Quality vs. Scale Trade-off: Attempting to scale output using generic templates destroyed reply rates. Creators routinely ignore template pitches that fail to reference specific, nuanced context from their actual video content.
- High Operational Cost: Agencies had to employ full-time Virtual Assistants or junior marketers strictly for data collection and manual drafting.
IMA (Influencer Marketing Agent) was engineered to solve this tradeoff between hyper-personalization and high-volume execution by converting the human agency workflow into a deterministic, autonomous agentic pipeline.
Architecture & How It Works
IMA utilizes a Client-First Multi-Agent Architecture running on top of Hermes Agent. The core execution pipeline relies on an orchestrated 5-agent fleet, where each specialist agent executes a tight, single-responsibility step within the pipeline.
[ YouTube Data API ] ──> ( 1. Discovery Agent )
│
▼
[ Firecrawl API ] ──> ( 2. Scraper & Transcript Agent )
│
▼
( 3. Context & Hook Extraction Agent )
│
▼
( 4. Copywriter Agent )
│
▼
( 5. QC & Dispatch Agent ) ──> [ Email Inbox ]
- 01
Discovery Agent
Interfaces with the YouTube Data API to filter channels by subscriber range, video upload frequency, niche keywords, and engagement metrics.
- 02
Scraper & Transcript Agent
Calls Firecrawl and targeted web scrapers to fetch raw video transcripts, channel metadata, contact emails, and social links.
- 03
Context & Hook Extraction Agent
Parses raw transcripts to extract core video thesis statements, timestamped topics, creator tone, and specific quotes suitable for dynamic hook generation.
- 04
Copywriter Agent
Combines creator context with predefined client offers to write personalized outreach copy, matching tone and highlighting specific moments from recent videos.
- 05
Quality Control & Dispatch Agent
Runs verification against anti-spam guidelines, checks for duplicate contacts, validates syntax, and queues emails for delivery.
Hermes Self-Learning Loop
IMA leverages Hermes Agent's continuous skill optimization. When an outreach sequence scores low engagement or encounters execution fallbacks (e.g., missing transcripts or API rate limits), the agent logs failure patterns, self-corrects parameter selection, and refines downstream extraction prompts dynamically without requiring code re-deployments.
Key Features
-
35+ Pre-built Agency Workflows
Modular YAML-driven execution routes for creator discovery, vetting, offer matching, follow-ups, and negotiation.
-
8 Specialist Agent Personas
Configurable roles ranging from Strategic Brand Lead to Copywriter and Data Analyst.
-
Transcript-Based Hook Extraction
Deep contextual analysis via Firecrawl to reference exact video moments rather than channel surface titles.
-
"Party Mode" Multi-Agent Brainstorming
Interactive session mode where specialist agents debate outreach angles and campaign positioning before output generation.
-
API Fallback Orchestration
Custom Python fallback mechanisms that seamlessly switch between primary scrapers, YouTube APIs, and alternative endpoints to guarantee 99.9% pipeline uptime.
-
Custom Email Persona Matching
Dynamically aligns outbound tone (casual, professional, technical) based on the target creator's video style.
Results & Metrics
Sourcing and dispatching to ~30 creators per inbox per day, running fully autonomously in production for internal agency operations.
Reduced campaign research and drafting time from 15–20 minutes per creator down to under 45 seconds of automated background processing.
Generated active interest and adoption from 50+ agency owners through live workflow demonstrations.
Achieved a 100% unique context match rate on outbound emails by embedding specific transcript references into every opening hook.
Tech Stack Deep Dive
- Python 3.13
- Primary language for orchestrator logic, async data execution, and custom tool bindings.
- Hermes Agent
- Core agentic execution runtime enabling self-learning loops, memory management, and tool-use lifecycle management.
- Firecrawl
- API used for deep web extraction and fetching structured video transcripts from dynamically loaded pages.
- YouTube Data API
- Upstream ingestion source for querying target channels, playlist items, and channel-level statistics.
- OpenCode
- Agent development harness and environment used for iteration, testing, and profiling.
- YAML
- Structured configuration format defining agent personas, prompt schemas, and multi-step pipeline routes.
Links & Resources
Lessons Learned
- 01
Granular Agents Outperform Monolithic Prompts
Attempting to have a single prompt discover, extract, and write email drafts consistently resulted in context drift. Separating execution into 5 single-responsibility agents dramatically improved output quality.
- 02
Scraping Fallbacks Are Critical
Relying on a single API or web scraper for transcript extraction causes pipelines to break at scale. Building multi-tiered fallback paths in Python was essential for maintaining high daily operational volume.
- 03
What I'd Do Differently
I would implement a strict human-in-the-loop (HITL) review dashboard earlier in the development process for batch approval before automated email dispatch, offering non-technical operators greater confidence during initial deployment phases.