Source

We created Source whilst working with MHI Media, a Bristol-based performance marketing agency scaling brands to 8 figures. Working with brands across fashion, lifestyle and wellness, MHI were already running hundreds of creator campaigns a year using across gifted seeding, affiliate campaigns and influencer deals. Creator campaigns are incredibly effective, but one admin can end up managing hundreds of different message threads across multiple platforms, with varying briefs, outreach campaigns, budgets and products to send. We set out to firstly solve the problem for them in house, and then to develop a self-service product that could be used by other brands and agencies.

The build lasted around 6 months, working across product design, system architecture, and full-stack engineering. From day one, our aim with Source wasn’t just to automate a workflow, but to build a single point of accountability and knowledge for a brand’s creator campaigns. Onboarding starts with a natural conversation so agent and brand team are aligned, and then Source develops a brand memory context that is then used throughout every stage of multi-week campaigns.

The hard part for a product like Source isn't conversation; it's building an agent that acts reliably over days, without supervision and across multiple concurrent tasks. We approached it as a team of specialised agents that hand work to one another, with a central coordinator that has memory, workspace context and pauses for a human's sign-off at the moments that matter.

With brand representation always a top priority, the system design had to provide the right amount of safety whilst automating the manual work. Making sure that a user can feel comfortable automating thousands of conversations meant building the right kind of guardrails for sending, reviewing and checking in when required. Campaigns in Source run as full simulations by default, allowing users to sandbox every interaction before going live. Safe mode, validation loops, and cross-campaign deduplication services all provide additional modes of security for users.

Enabling agentic decision making in an affiliate campaign context means first agreeing on sources of truth and knowing when to escalate things to review. Source treats outreach as a timeline of steps (pitching, follow-ups, negotiation, logistics, and closing), each with its own validation criteria that can be assessed by a team of sub-agents before proceeding. At optional ‘brand review’ checkpoints the system pauses for human sign-off or input, for instance to add a contract URL or confirm a payment. Source generates a plan for each campaign dynamically, runs it in simulation with creators, and then amends the timeline based on user feedback before anything goes live.

For platform-wide search and critical analysis of creators, we leveraged several parts of the Parallel Web Systems API within a multi-agent architecture that had access to the Influencers Club database. Influencers Club gave us API access to millions of creator profiles across every platform, whilst the agent harness meant that those profiles could actually be reflected upon against the brand.

In a multi-agent architecture like this, one agent can query creators from the database whilst another analyses them against dynamic conditions informed by brand memory and campaign goals. As results come in and fit is assessed, another agent can assess and refine the base queries based upon analysis. This allows the system to continuously spin up searches and channel creators into active campaigns whilst incorporating user feedback and optimising the next cycle.

On top of this, we engineered our own OpenClaw-style harness to relay messages and embed Source directly in teams via. Slack, WhatsApp and Email. We also automated the unglamorous part: the email setup and deliverability that normally takes a team to manage, so a brand can scale to sending hundreds of emails a day without touching any infrastructure.

In the first 4 weeks of beta testing, Source onboarded 63 external brands and helped teams discover 22,361 creators across platforms, enriching 4,286 with verified email. In that window, brands sent 3,274 real emails (excluding simulation) using Source and launched 48 end-to-end campaigns that are still being managed autonomously.

One of the highlights for us has been seeing brands transition away from the web interface and into Slack, just like working with an expert colleague. New creator cohorts can be found and contacted with a couple of back and forth messages, with updates dropped asynchronously by Source when they come in.

Designing and iterating upon Source proved to us the power of combining specialised, long-running agents into a harness that can reliably coordinate the hundreds of other small, critical decisions that occur in real-world tasks. In order to collaborate over weeks and months, context and coordination have as much a part to play as raw intelligence.