How Qutron Agentic AI works
Qutron Agentic AI
Nine specialist agents draft the plan, the forecast and the deck from your own data. You review and approve before anything reaches a client.
The first draft is where agency weeks disappear
- Three days pulling numbers before anyone writes a strategy
- Benchmarks living in one analyst's spreadsheet
- Decks rebuilt from scratch for every pitch
- No record of why a recommendation was made
- A brief becomes a costed draft plan, forecast and deck
- Every number sourced from your own connected data
- Proven structures reused from decks you already won with
- Every prompt, model call and decision traced and replayable
Built so a failed run never costs you a deadline
Durable workflows, traced decisions, no black boxes.
Inngest-orchestrated workflows
Every brief becomes a durable workflow. Each agent runs as an isolated Inngest step with automatic retries, exponential backoff, and exactly-once semantics. A failed model call or transient API blip never loses your run.
- Per-step retries and dead-letter handling
- Resumable runs survive deploys and restarts
- Replay any historical run from the dashboard
Langfuse traces every decision
Every prompt, model call, tool invocation, latency, and token cost is traced end-to-end. You can audit exactly why an agent recommended what it did, replay it, and score it for quality.
- Per-run trace tree with prompt + completion
- Cost and latency budgets per agent
- Quality scoring and prompt regression catches
One brain. Many specialists.
A central brain orchestrator routes the brief into specialized branches: deck, media plan, SOW, email, and package. Each branch composes the right agents for that artifact, in parallel where possible.
- Per-artifact branch (deck / plan / SOW / email)
- Specialist agents share a typed contract
- Composable: add a new branch without rewiring
Friendly errors, automatic recovery
Failures are sanitised into human messages and surfaced with a reference ID. Stuck runs are auto-resumed; partial outputs are preserved so a re-run never starts from zero.
- Automatic resume of stuck synthesis
- Reference ID on every user-facing error
- Full technical detail kept server-side
Nine specialists, three waves, one draft
Data agents gather, analysis agents reason, one composes the output.
Wave 0: Data Sources
Independent agents that gather raw intelligence from plans, benchmarks, history, and audiences.
Planning Agent
The budget architect.
Pulls budgets, channel mix and line items from your live media plans.
Benchmark Agent
Industry truth, quantified.
Brings channel benchmarks by market and vertical into every recommendation.
Historical Agent
Lessons from every campaign.
Reads past campaign performance for trend and seasonality context.
Insight Agent
Audiences revealed.
Surfaces audience research and demographic signals behind the strategy.
Wave 1: Analysis Layer
Reasoning agents that synthesize Wave 0 outputs into strategy, forecasts, and recommendations.
Strategy Agent
From data to direction.
Turns the gathered data into recommendations, risks and opportunities.
Forecast Agent
Tomorrow's metrics, today.
Projects KPIs and return across scenarios before budget is committed.
Media Agent
Right channel. Right placement.
Recommends channel and placement mix against reach, frequency and cost.
Wisdom Agent
Your best work, remembered.
Reuses proven structures from your own past decks.
Wave 2: Visual Composition
The final layer that transforms all resolved data into presentation-ready layouts and narratives.
Makeup Agent
Data, beautifully composed.
Composes the final layouts, charts and slide order.
Follow one brief through the mesh
Agents run in parallel where they can, in sequence where they must.
"Q3 launch across MENA. Budget: $2.4M. Objective: awareness + consideration."
Dependency graph executor
32 slides generated with strategy validated against benchmarks. Layouts optimized for executive audience.
Six rules the agents never break
Durable by Default
Every run is an Inngest workflow. Steps retry on transient failures, resume after deploys, and can be replayed from the dashboard.
Traced End-to-End
Langfuse captures every prompt, completion, tool call, latency, and token cost so you can audit, score, and improve every recommendation.
Parallel Execution
Independent agents run as concurrent Inngest steps, not a sequential chain, so the time-to-draft is set by the slowest agent, not the sum of all of them.
Orchestrated Workflow
A central brain orchestrator routes the brief into the right branch (deck, plan, SOW, email) and only dispatches each agent once its prerequisites are satisfied.
Collective Intelligence
No single agent has the full picture. Every output is the synthesis of multiple specialists, grounded in your benchmarks and historical performance.
Human-in-the-Loop
AI proposes. Your team disposes. Every generated output passes through configurable review stages before it reaches a client.
Brief in, reviewed artifact out
Bring one live brief and watch it drafted
We connect a data source, run the mesh on your brief and show you the trace behind every number.