← back to davelee.ai
CASE STUDY · 02 AGENTIC OS · 2025–PRESENT

Navos — an agentic OS for marketing operations.

A platform that connects creative, optimization, assets, wallet, and orchestration. The user no longer drives a tool — they delegate a goal, and a coordinated set of agents runs the loop.

Role
AI Product Director
Org
Tec-Do
Surface
Agent platform + console
Audience
Performance marketers · in-house growth

§01 Premise

Most "AI marketing" products are single-shot tools attached to a chat box. They generate, then stop. Real growth work is continuous: a goal sits at the top, dozens of decisions hang below it, and feedback rewrites the plan every day.

Navos treats the goal as a first-class object. A planner decomposes it, specialist agents execute, a wallet polices spend, and a memory layer captures what worked.

§02 Topology

PLANNER goal · plan · route CREATIVE brief · gen · variants OPTIMIZER spend · audience · creative ASSETS memory · DAM · brand WALLET budget · policy · risk ORCHESTRATOR queues · retries · audit

§03 Run trace

What the user sees when a goal is dispatched. The trace is also the audit log; nothing happens off-record.

t+0.0s PLANNER decompose("grow EU signups, $50k cap, 14d") ✓
t+0.3s WALLET cap=$50,000 · daily=$3,500 · risk=normal ✓
t+0.6s ASSETS load brand v3 · tone · banned terms · winners(30d) ✓
t+1.2s CREATIVEgen 24 variants · 3 markets · 4 channels ✓
t+18s ORCH handoff to platforms · stage · review queue ✓
d+1 OPTIMIZERkill 6 underperformers · scale 2 winners +30% ✓
d+3 ASSETS writeback: motif "outdoor + handheld" → winners ✓

§04 Design principles

// goal as object

  • Goals are persistent, not prompts
  • Plans are revisable, not one-shot
  • Every agent action ties back to a goal

// trust before autonomy

  • Wallet caps before launch
  • Policy gate on every external action
  • Auditable trace, not a hidden chain

// memory as moat

  • Brand, motifs, winning DNA persist
  • Writeback after every campaign
  • Memory is what compounds, not models

// substrate-agnostic

  • Models behind a stable runtime
  • Channels behind a stable adapter
  • The grammar is the durable surface

§05 What I did

// product

  • Architected the planner / agent / wallet model
  • Defined the trace + audit experience
  • Memory schema and writeback semantics

// strategy

  • Positioning vs. single-shot tools
  • Migration story from Tec-Creative customers
  • Channel partnership scoping
← back to davelee.ai