Commercial · Google Ads SEM campaign steering

sem-agentic

Google Ads budget runs efficiently only when campaign steering and impact measurement work together. Augur is the working proof for the steering side.

Google SEM · workflow in 5 steps

1. Keyword research

2. Campaign structure

3. Ad copy

4. Bidding strategy

5. Optimisation & analysis

An agentic system that analyses this workflow and recommends optimisations.

Scale Maintain Investigate Pause

Objective

Google Ads budget delivers only when campaign steering and impact measurement work together.

Augur is the working proof for the steering side: a multi-agent system that maps keyword strategy, performance analysis and budget-relevant recommendations for a Google SEM campaign end to end — with the same measurement discipline that signal-steering demonstrates on the evaluation side.

Approach

Five validated agents, a coordinator with self-checking and a staged rollout.

Each agent produces reasoned proposals and is checked in a way suited to its output: shadow log and human review, LLM-as-a-judge or hypothesis tests. A strategy coordinator consolidates the results into a traceable recommendation. Deterministic logic where the rules are clear; an LLM only where it adds value. Execution progresses in stages: Shadow → Sandbox → Limited Live → Autonomous.

Build vs. Buy

Introduce risk controls before autonomy instead of buying blindly or building without validation.

The simulation/production mode switch and five-stage rollout model (shadow analysis → shadow implementation → sandbox → limited live activation → autonomous operation) are the architectural answer to this question. Stages 1–2 are implemented; 3–5 are documented extension points, not live operation. Which proposals are adopted and from which stage an agent may act autonomously remains a deliberate human decision.

Evidence-backed

Agentic SEM system as a running pilot
Validation per agent and coordinated consolidation
Documented evaluation (ADRs, evaluation framework, audit)

Not evidenced · out of scope

Autonomous live operation
Impact at scale
Performance foundation partly synthetic
Execution in early rollout stages under human control

Augur operator dashboard, the Streamlit interface of a complete coordinator run: KPI cards, pipeline status, recall history, critique results and prioritised cluster actions.

Augur operator dashboard with KPI cards, pipeline status, recall history, critique results and prioritised cluster actions

Augur · Architecture

Multi-agent system with staged validation

Five specialised agents produce reasoned proposals. Each agent passes through a validation layer suited to its output. The strategy coordinator consolidates the results, with self-checking, into an auditable recommendation.

Brief NordicGlow Skincare D2C, DACH market entry

Research-Agent

Keyword discovery from market data

ToolDataForSEO API

PatternReAct with tool loop

CheckShadow Log + Human

Cluster-Agent

Semantic keyword clustering

ToolEmbedding + pgvector

PatternVector-Similarity

CheckLLM-as-a-Judge

Performance-Analyst

Data analysis, hypothesis tests

ToolSQL + statistics libraries

PatternHypothesis-Testing

CheckHistorical backtest

Copy-Agent

Ad copy with brand voice

ToolLLM + Brand-Prompt

PatternConstrained Generation

CheckJudge + Human-Review

Strategy-Coordinator

Consolidates the four agent outputs into a coherent strategy recommendation. Performs a self-check, identifies contradictions, requests revision and documents confidence and open risks.

Self-Reflection-Pattern Python + LangGraph Confidence-Scoring Audit-Trail

Proposal Cluster + budget + copy + risks, structured and auditable

Architecture principle: proposal generation, not action execution. Execution is a separate layer with a staged rollout (Shadow → Sandbox → Limited Live → Autonomous).

Data foundation

PostgreSQL + pgvector

Live DataForSEO data

Synthetic performance data (90 days)

n8n workflow orchestration

Tool What the agent calls

Pattern Reasoning strategy

Check Validation method

Build versus buy as an architecture decision

The simulation/production mode switch and five-stage rollout model (shadow analysis → shadow implementation → sandbox test → limited live activation → autonomous operation) are not incidental. They are the architectural answer to a leadership question: when should you buy a finished solution, when should you build, and how much autonomy must a system earn before moving real budget? Stages 1–2 are implemented (the agent analyses, a human reviews); stages 3–5 are documented extension points, not live operation. This risk-before-autonomy logic is transferable, whether Augur itself or a third-party solution is used in the end.

Product characteristics: what differentiates it

Two-layer architecture (ADR-015)

Deterministic action layer (pre-flight rules) plus optional LLM analyst layer. Layer 2 remains subject to evidence and is not implemented until its value is demonstrated.

Two-Track Eval (ADR-013)

F1 versus domain ground truth plus rule compliance as orthogonal dimensions. Evidence: the LLM meets 6% of constraints by itself; deterministic logic secures 94%.

Cluster aggregation against prompt collapse (ADR-014)

626 keyword-forced recommendations → 17 cluster requirements, prompt reduced from 33k to 3–4k tokens. Zero determinism violations across N runs.

Selectable LLM backend per step

Sonnet for reasoning, Ollama for low-cost calls, configurable via llm_overrides.

Mode-Switch Simulation / Production

Mock database for demo and development, live DataForSEO for production, with no code change required.

Variante-B: n8n-Wrapper → Python-Endpoint (ADR-012)

Single source of truth in the Python coordinator, n8n as an HTTP proxy, replicated three times (cluster-full, strategy, reflect).

Hybrid keyword generation

Anchor clusters as the skeleton plus parallel LLM variation per anchor, producing 401 keywords with a realistic long-tail distribution.

Best practices as a YAML library

Declarative rules (for example v2026_05.yaml), versioned, with custom overrides available.

Sources: 15 ADRs · reflections · status.md · code patterns · audit.md · Maturity-v3

sem-agentic · architecture proof

Augur

A multi-agent system for SEM performance. Deterministic where the rules are clear. An LLM only where it adds value.

Brand brief + keyword set

Input to a complete coordinator run

Sourcing & clustering n8n · Ollama · pgvector

Research

Keywords via DataForSEO

Embed

Vectors, nomic-embed

Cluster

Assign + LLM-Judge

Analysis & strategy LangGraph-Coordinator · Python

Performance

Hypothesis testing + rules

Copy

Ad copy per cluster

Synthesis

Recommendation generation

Reflect

Self-check

↺ Refine loop for unresolved issues · max. 2 iterations

The core: recommendation generation

The architectural boundary runs here: not between components, but within the decision.

Layer 1 · deterministic

Pre-flight rules → forced actions

A versioned best-practice rule set evaluates the data snapshot and enforces the cluster action. Auditable, reproducible, with no model involved.

developed · active

94 %

of system rules
secured deterministically

— vs —

6 %

met by the 7B model
on its own

Layer 2 · LLM reasoning

Prioritisation & cross-cluster narrative

Intended to contribute judgement where no rule applies. Today it contributes rationale only; independent reasoning remains subject to evidence.

direction decided · not implemented

Prioritised cluster actions

Scale · Maintain · Investigate · Pause, with rationale, confidence, impact & full lineage

What differentiates the architecture

Deterministic where rules apply

The action comes from the rules, not from the LLM. The model writes the rationale; it does not make the decision.

Mode-Switch Simulation ↔ Production

Mock database for zero-cost demos via replay cache; live DataForSEO for production. Switching requires no code change; simulation is the default.

Fact-checking as a correctness anchor

An evidence validator checks every LLM figure against the database. It exposed where the model hallucinated causality and set the reasoning proof to an honest FAIL instead of a polished pass.

Stack

LangGraphn8nFastAPIPydantic v2PostgreSQL + pgvectorStreamlitDocker

Models

qwen2.5:7b · Default nomic-embed Sonnet · opt-in reasoning

Verified architecture state · 9 workflows · 7 coordinator stages · 17 clusters

What this means for a marketing organisation

The performance analyst agent uses hypothesis testing against backtesting data — the same logic as incrementality-led budget steering, but at campaign rather than channel level. For a marketing organisation, this means Google Ads campaigns can be steered in a structured, auditable way before an agent is ever allowed to decide autonomously. The system can generate proposals. Which proposal is adopted, and how much autonomy it earns, remains a guided human decision.

How I work with AI tools as a leader: AI engineering practice →