Case study · Tec-Do · 2025–now

Navos: a multi-agent platform for global marketing

Navos takes a marketing goal and runs it with a team of specialist agents: market analysis, creative generation, launch, and optimization in one loop, with every step on the record.

Head of Navos·2025–·navosagent.ai ↗

LayersResearchModelSystemProductGrowthRevenue

At a glance

$100M+annual ad spend running through Navos

months → hoursa marketing cycle that took one to three months now runs in hours, per core customersSource: 智东西 / 36Kr, 2026-03-04 ↗

< 5 minfrom “find UK/US pool-robot influencers” to five outreach emails sent, watched live at WAIC 2026Source: 数据猿, 2026-07-21 ↗

01context

Context

Most AI marketing tools are single-shot: they generate an asset and stop. Growth work is continuous. A goal sits at the top, dozens of decisions hang below it, and feedback rewrites the plan every day.

A marketing cycle that used to take one to three months had to become something a team of agents could run in hours, without giving up control of budget and brand.

02system

How it works

Navos treats the goal as a first-class object. A planner breaks it down, specialist agents do the work, a wallet enforces budget and policy before anything launches, and a memory layer keeps what worked.

Simplified architecture.

The first release (2025) shipped four modules: market insight, creative radar, ad monitoring, and data reporting. Navos 2.0, launched at WAIC in July 2026, chains market analysis, creative generation, and ad delivery into one multi-agent workflow.

Design principles

01Goals are objects, not prompts

Goals persist and plans get revised. Every agent action traces back to a goal.

02Trust before autonomy

Budget caps before launch, a policy gate on every external action, and an auditable trace instead of a hidden chain.

03Memory is the moat

Brand rules, motifs, and winning creative persist and get written back after every campaign. Memory compounds; models don't.

04Models are swappable

Models sit behind a stable runtime and channels behind stable adapters, so the workflow outlives any one model.

03role

What I owned

Architecture: the planner / agent / wallet model, the audit trace, and the memory schema.
Roadmap and evaluation: what ships next, and the evaluation framework that decides when an agent is ready for production.
Agent infrastructure: started and led the Marketing Agent Infrastructure program, the shared platform, tooling, and engineering standards every agent is built on.
Team: built and led the team that ships Navos, and incubated the Social Media Agent.

04more

Coverage & related

Coverage

→ Note: the logic of next-gen growth systems → Paper: Fast Models, Slow Evidence (arXiv, 2026)→ Note: local LLMs and the right harness