Sixteen agents, one objective: your profit.

The marketing line of every PPC tool built since 2022 is "AI-powered." The architecture behind that line, in most cases, is a single optimisation engine — a gradient-boosted model, or a reinforcement learner, or a rules table with a learned-weight layer on top — wearing the label AI. One model, one objective, one budget allocator.

Mirox is built differently. The intelligence pillar — the third of four — is sixteen specialised agents, each with its own scope, its own reasoning, and its own decision trace. They coordinate on a shared data structure called the blackboard, which is how a team of specialists actually thinks about a single account.

This post is the long argument for why the architecture matters and a tour of the four named agents you will see in every trace.

"AI" without an org chart

A single optimisation engine works fine in a single-objective problem. Maximise clicks per dollar. Minimise ACoS. Hit a TACoS target. The single objective is encoded in the loss function, the engine optimises it, the seller gets a number.

Amazon advertising is not a single-objective problem. It is the simultaneous solution of:

• Bid-level economics (pCVR × margin × placement modifier).
• Search-term relevance (does the query match the product, semantically).
• Inventory constraints (do not bid on what you cannot ship).
• Stage-aware strategy (a launch campaign should not be optimised like a mature one).
• Cross-campaign cannibalisation (don't let two campaigns bid against themselves).
• Safety constraints (the seven gates from the safety post).
• Marketplace specificity (German buyer behaviour is not English buyer behaviour).
• Long-tail keyword discovery and harvesting.

A single engine that tries to encode all of this in one loss function ends up either too vague to outperform a thoughtful human or too brittle to survive an Amazon catalogue update. The category answer has been to ignore most of the list and optimise hard on one or two — usually bid-level ACoS — and ship the rest as configurable rules.

The agent answer is to give each part of the problem its own specialist with its own reasoning. The blackboard architecture is forty years old in AI research; the surprise is not that it works, but that nobody else applied it to PPC.

The four named agents

The full system has sixteen agents. Four are named in every external trace because they are the four a seller meaningfully interacts with. The others — among them Executor, Stress-Tester, Harvester, Blackboard Coordinator — are infrastructure.

Tactician — the bid proposer

Tactician is the agent that proposes a bid for a keyword at a moment in time. The math underneath is Thompson sampling over a BetaBinomial pCVR posterior, multiplied by an asymmetric semantic relevance score, multiplied by the supply multiplier from Sentinel, then walked through a candidate-set selection to pick the bid that maximises expected profit given the constraints in force.

What that means in practice: Tactician is the agent making the most micro-decisions. It writes the largest share of the trace surface. The pCVR estimate, the relevance score, the alternatives ruled out — all Tactician. If you read one trace per day, it will be a Tactician decision.

Tactician is also the agent with the most narrowly-scoped competence. It is not strategic. It does not decide whether a campaign should exist. It bids on the campaigns and keywords that exist, given the inputs it has. The strategic context is supplied by Strategist.

Strategist — the stage and context agent

Strategist is the agent that decides what stage a campaign is in and what bidding philosophy should apply at that stage. A new campaign with no conversion data is in Learning and should bid widely to harvest signal. A mature campaign with a strong pCVR is in Optimising and should bid narrowly on what works. A campaign post-launch trailing off is in Maintenance and should be allowed to consolidate.

These transitions are not configured by the seller; they are inferred by the agent from the campaign's own performance trajectory. The state machine is finite, the transitions are reversible, and every transition writes to the trace with the trigger that fired.

Strategist's effect on a single bid is indirect — it sets the philosophy Tactician is optimising inside. But on the scale of a campaign quarter, the strategist decisions move more spend than the tactician decisions, because a wrong stage causes every bid inside it to be wrong.

Sentinel — the inventory watch

Sentinel was promoted to a fourth supporting pillar under Safety in 2026 Q1. The full treatment is on the safety post. The short version: Sentinel reads days-of-cover from the SP-API inventory feed and applies a supply multiplier — sigmoid in days-of-cover — to every bid Tactician proposes. It is the agent that prevents the AI from accelerating PPC spend into a stockout.

Sentinel is the one agent that is not adversarial to the AI. It exists to protect against a physical constraint the bidding logic cannot see.

Semantic — relevance and search-term harvesting

Semantic is the agent that computes how well a search term matches a product detail page. The underlying model is asymmetric — the query embedding uses a RETRIEVAL_QUERY task and the document embedding uses a RETRIEVAL_DOCUMENT task — which is the right modelling choice for a query-to-document similarity problem and is consistently better than the symmetric default in our internal evals.

Semantic is also the harvester. It watches the search-term reports, identifies high-relevance terms that are currently auto-targeted, and proposes them for promotion to explicit keyword targets. The reverse, too — when an explicit keyword is auto-bidding into low-relevance traffic, Semantic flags it for review.

The trace surface for Semantic is the score, the embedding model version, and the relevance gate. Sellers can audit which terms cleared the gate and which were rejected.

The remaining twelve

Twelve agents do not appear by name in the typical trace because they are infrastructure rather than decision-makers — Executor (the SP-API write path), Stress-Tester (the offline replay of the day's decisions), Harvester (search-term mining), Blackboard Coordinator (the agent that resolves conflicts when two agents propose competing actions), and a handful of monitors.

They will get their own page when we ship /agents in 2026 Q3. Until then, the four above are the ones to know.

How the agents coordinate

The architecture is a blackboard. Each agent reads from a shared state representation, writes its proposals back to the same state, and the Coordinator resolves conflicts using a pre-declared priority order.

Concretely: Tactician writes a proposed bid. Sentinel writes a supply multiplier. Shield writes a safety gate state. The Coordinator computes the final bid as (Tactician's bid) × (Sentinel's multiplier), subject to Shield's gates passing. If two agents propose conflicting non-multiplicative actions — say, Strategist proposing a stage transition that contradicts Sentinel's rationing flag — the Coordinator applies the priority order and writes both proposals plus the resolution to the trace.

The blackboard is what makes the architecture transparent. There is no hidden coordination — the agents are not "talking" to each other in some opaque message bus. They are reading and writing a state object that the seller can read in the dashboard.

Real models, not rules dressed up

A line we use internally: real models, not rules dressed up as AI. The category does the inverse — rules dressed up as AI — often enough that it is worth being concrete about what real models means.

• The pCVR estimator is a hierarchical BetaBinomial. The hierarchy is per marketplace, per category. The prior is fit weekly from aggregated data; the posterior updates per impression. This is the same modelling shape that wins on conversion-rate problems in advertising research papers; we did not invent it but we are one of the few PPC products that actually ships it.
• The semantic embedding is the asymmetric variant of a current open-weights model, hosted on EU infrastructure. The asymmetric variant outperforms the symmetric default by roughly 19 percent on internal relevance evals.
• The bid candidate selection is Thompson sampling over the posterior, not greedy maximisation. Thompson sampling explores the bid space proportionally to the posterior uncertainty, which is the right exploration policy for a non-stationary problem. The greedy alternative under-explores in low-data regimes and over-commits in high-data ones.
• The state machine in Strategist is reversible by construction. Reversibility matters because a one-way state machine — common in rules engines — accumulates state drift that requires manual reset. A reversible one self-heals.

None of the above is novel research. The combination, shipped in a single coherent product, is rarer than it should be.

What you see in the trace

The agent architecture matters because it is visible. Every trace names the agents involved. Every decision attributes itself to the agent that proposed it. Every conflict resolution names the agents that conflicted and the priority that resolved them.

Open the dashboard. Click into a bid. Read the trace. There is no "the model decided." There is Tactician proposed €1.84 against alternatives of €1.62 and €2.10, Sentinel applied a supply multiplier of 0.85, Shield's ROAS gate held at the 0.62 threshold, Coordinator resolved the final bid at €1.56.

That is the difference between a tool that markets as AI and a tool that is a team.

Where this pillar leads

Intelligence without transparency is a black box claiming to be smart. Intelligence without safety is a fast bidder with no brakes. Intelligence without aligned pricing is an expensive engine for someone else's commission.

The fourth pillar — aligned pricing — closes the loop on why the architecture serves the seller and not the vendor. Read why we will never charge a percentage of your ad spend next.