The AP3 agent system, in section.

Concept diagram of how AP3 works. It is not a client system, a client result or a forecast.

The whole system

0102030405060708091011121314151617181920A

L6 / Gap

  1. 20Launch, operate, measureProduction, launch and measurement have no skill yet. This is the gap.Gap · Not built
  2. AA person decidesEvery gate is held by a person: scope, plan and acceptance are not delegated.Person · Done by a person

L5 / Build and verify

  1. 16Goal runnerTakes a goal to a verified finish: plan, build, verify, repeat up to three cycles.Skill · BuiltDetail view
  2. 17Scope and criteriaFixes scope, numbered acceptance criteria and the definition of done.Skill · Built
  3. 18BuilderBuilds the artifact and checks its own work.Skill · Built
  4. 19Independent verifierA cold agent that did not build the work. Returns a verdict per criterion, with evidence.Cold agent · Built

L4 / Rehearse with the buyer

  1. 10Decision-group rehearsalRuns a buyer's decision group through several rounds and returns hypotheses to test.Skill · Smoke-tested onlyDetail view
  2. 11Stakeholder agentsOne cold agent per stakeholder, one turn per call.Cold agent · Smoke-tested only
  3. 12Buyer-side offer checkOne cold verdict from the buyer's side: would they buy at this price.Cold agent · Built
  4. 13Audience rehearsalRuns a consumer ad campaign through a synthetic social world over several days.Skill · Smoke-tested onlyDetail view
  5. 14Consumer agentsOne cold agent per consumer, one round per call.Cold agent · Smoke-tested only
  6. 15Front-desk agentAnswers messages within a written policy and nothing else.Cold agent · Smoke-tested only

L3 / Plan

  1. 08Campaign planningTurns an approved recommendation into a campaign strategy brief. Plans only; builds and launches nothing.Skill · Built
  2. 09Marketing leadPositioning and demand strategy. Convenes a specialist for bounded work.Skill · Not yet runDetail view

L2 / Understand and diagnose

  1. 05Pre-call researchResearches a company from public sources before a call and writes a brief.Skill · Built
  2. 06Problem diagnosisWorks out what the organisation should do and why, then recommends and hands off.Skill · BuiltDetail view
  3. 07Venture reviewNames the constraint on a venture and decides whether to think, test or build.Skill · Built

L1 / State

  1. 01Discovery briefsResearch briefs written before a discovery call.State · Built
  2. 02Case files and run logsWorking folders for diagnosis, planning and goal runs.State · Built
  3. 03Simulation recordsTranscripts and scorecards from rehearsals.State · Built
  4. 04Project stateOne record of what each project is and where it stands.State · Built
All relationships, as a list
  • Problem diagnosis hands off to Campaign planning.
  • Venture review optionally feeds Marketing lead.
  • Campaign planning optionally feeds Decision-group rehearsal.
  • Builder hands off to Independent verifier.
  • Independent verifier returns to Builder.
  • Independent verifier optionally feeds Launch, operate, measure.
  • Skill: runs in the working session
  • Cold agent: isolated, sees one packet
  • Dashed outline: conditional, done by a person, or not built
  • Red bar: the one gate. It marks the gap
  • Solid line: hand-off. Short dashes: returns. Long dashes: optional
  • Black mast: a person holds every gate
Status key
Status vocabulary
BuiltThe skill or agent definition exists and is installed. This page does not claim it has run for a client.
Smoke-tested onlyRun end to end on a small test case to confirm it works. Not used on real engagements.
Not yet runWritten and reviewed. Its behaviour tests were read, not executed against a model.
Not builtSpecified, not produced.
Done by a personA decision or step that stays with a person and is not software.
Conditional pathA path that exists only when a stated condition is met.

The gap. Launch, operate and measure are not built. Agents research and check; a person scopes, decides and accepts.

Detail views

Detail of part 10 · Decision-group rehearsal

Decision-group rehearsal, in section

121314151011060708090501020304C1

L5 / Synthesis

  1. 01ReportClaims, grounding and leakage checks.Skill · Smoke-tested only
  2. 02Claim classesClaims sorted by their relation to the baseline; class C is the finding.Skill · Smoke-tested only
  3. 03Independent verifierOptional cold check of the scoring.Cold agent · Smoke-tested only
  4. 04HypothesesTo test in real conversations.Skill · Smoke-tested only
  5. C1Never a predictionForecast framing is corrected out of the report.Gate · Built

L4 / Run 2, identical inputs

  1. 05Second independent runSame inputs. Only recurring hypotheses count.Skill · Smoke-tested only

L3 / Run 1, in parallel

  1. 06First reactionsRound 1: independent, no earlier round read.Cold agent · Smoke-tested only
  2. 07ResponsesRounds 2 and 3: each reads every earlier turn.Cold agent · Smoke-tested only
  3. 08InterviewThe decision question, asked after all rounds.Cold agent · Smoke-tested only
  4. 09Stakeholder agentsSix cold agents, one turn per call, four turns each.Cold agent · Smoke-tested only

L2 / Baseline, before any run

  1. 10BaselineWritten before any run, from the evidence and every brief.Skill · Smoke-tested only
  2. 11Baseline fingerprintA hash that fixes the baseline before the agents start.Skill · Smoke-tested only

L1 / Evidence and briefs

  1. 12Evidence registerFacts and assumptions, tagged separately.State · Built
  2. 13Shared briefWhat the whole group has seen.State · Built
  3. 14Private briefsEach stakeholder agent sees only its own.State · Built
  4. 15Leakage markersA list used to check that private details did not leak.State · Built
Relationships, as a list
  • Evidence register hands off to Baseline.
  • Baseline optionally feeds Stakeholder agents.
  • Stakeholder agents hands off to Second independent run.
  • Second independent run hands off to Claim classes.

Detail of part 13 · Audience rehearsal

Audience rehearsal, in section

1314151617101112050607080901020304C1

L4 / Checks and synthesis

  1. 01ReportFriction points and validation questions.Skill · Smoke-tested only
  2. 02DiagnosticsDiagnostics, not predictions.Skill · Smoke-tested only
  3. 03Artifact checksA blinding flag voids attention claims.Skill · Smoke-tested only
  4. 04World recordThreads, chats and memory that the synthesis reads.Skill · Smoke-tested only
  5. C1No claim on attentionIf the check raises a blinding flag.Gate · Built

L3 / The round loop

  1. 05PacketsOne line per active persona.Skill · Smoke-tested only
  2. 06Consumer agentsOne round each, in parallel, cold and isolated.Cold agent · Smoke-tested only
  3. 07ApplyThe engine updates the world.Skill · Smoke-tested only
  4. 08Front-desk agentAnswers within the policy only.Cold agent · Smoke-tested only
  5. 09Apply repliesReplies land in the chats.Skill · Smoke-tested only

L2 / Engine, owns all state

  1. 10Simulation engineDecides who sees what and when. Owns attention.Skill · Smoke-tested only
  2. 11World statePersonas, threads, chats and memory.State · Built
  3. 12BaselineFixed before round zero.Skill · Smoke-tested only

L1 / Evidence pack

  1. 13Campaign settingsPopulation, targeting and attention settings.State · Built
  2. 14Ads under testThe ad plus two real competitor ads.State · Built
  3. 15Advertiser lookupsA quick and a deep read of every advertiser.State · Built
  4. 16Customer voiceReal customer phrases only.State · Built
  5. 17Front-desk policyEverything the front desk may say.State · Built
Relationships, as a list
  • Simulation engine hands off to Apply.
  • Apply hands off to World record.

Detail of part 09 · Marketing lead

The marketing lead and its specialists, in section

010203040506070809101112C1

L4 / Run modes

  1. 09InlineSame voice, headed pass, no independence.Skill · Not yet run
  2. 10SpawnAn isolated agent, run in parallel.Cold agent · Not yet run
  3. 11ExportThe owner pastes the brief into a fresh chat.Cold agent · Not yet run
  4. 12The brief, carriedEvery field required; 'unknown' if none.Note · Not yet run
  5. C1Own work never inlineCritique is spawn or export only.Gate · Built

L3 / The justification test

  1. 07Would it change the decision?Depth, critique, wide research or parallel work.Skill · Not yet run
  2. 08Specialist briefEleven fields. Only the relevant state.Skill · Not yet run

L2 / The lead's own work

  1. 04Lane checkNot marketing: say so and hand back.Skill · Not yet run
  2. 05Seven layers, in orderStop at the first that fails.Skill · Not yet run
  3. 06Think, test or buildThe leftmost route that fits.Skill · Not yet run

L1 / State

  1. 01Saved stateOnly the change since last time, read back.State · Not yet run
  2. 02Standing instructionsFound and followed first.State · Not yet run
  3. 03Manual save fallbackIf state cannot be saved, the owner stores the file.Gap · Done by a person
Relationships, as a list
  • Saved state hands off to Lane check.
  • Seven layers, in order optionally feeds Specialist brief.
  • Specialist brief hands off to Spawn.

Detail of part 16 · Goal runner

Goal runner and independent verifier, in section

010203C10405060708091011

L4 / The iteration gate

  1. 09Build and verify cyclesThree at most, then stop.Skill · Built
  2. 10A person acceptsBusiness acceptance is not the runner's.Gap · Done by a person
  3. 11Stops regardlessSpend, send, delete, access, commitments.Gap · Done by a person

L3 / Stage 3, cold

  1. 07Independent verifierA verdict per criterion, with evidence.Cold agent · Built
  2. 08Lesson captureOnly if a lesson emerged.Gap · Conditional path

L2 / Stages 1 to 2, inline

  1. 04Scope and criteriaScope, numbered criteria, definition of done.Skill · Built
  2. 05Acceptance criteriaFixed. Failures fix the artifact.Skill · Built
  3. 06BuilderBuilds and checks its own work.Skill · Built

L1 / Stage 0, the gate

  1. 01Outcome and observable criteriaOr a stop, never a guess.Skill · Built
  2. 02Named checkThe check that gives the evidence.Skill · Built
  3. 03Unrunnable goalReported as undefined, not interpreted.Gap · Conditional path
  4. C1No criteria, no runThe gate exists to stop.Gate · Built
Relationships, as a list
  • Outcome and observable criteria hands off to Scope and criteria.
  • Builder hands off to Independent verifier.
  • Independent verifier returns to Builder.
  • No criteria, no run guards Outcome and observable criteria.

Detail of part 06 · Problem diagnosis

Problem diagnosis and its workers, in section

010203040506070809101112C1

L4 / What the worker returns

  1. 10Brief ready or provisionalPath, short answer, gate results.Skill · Built
  2. 11Needs inputEach question, with a default.Skill · Built
  3. 12Revalidation requiredBack to the case. Never argued away.Gap · Conditional path

L3 / On a person's word only

  1. 07RecommendationPaid acquisition, with its readiness check.Skill · Built
  2. 08Cold workerA general-purpose agent in worker mode.Cold agent · Built
  3. 09Marketing leadOptional, beside it, no coordination.Gap · Conditional path

L2 / Entry readiness

  1. 04KnownGoal, constraint, rationale, target, limits and evidence.Skill · Built
  2. 05PartialResolve before handing off.Skill · Built
  3. 06UnknownSends the case straight back.Gap · Conditional path
  4. C1Nothing starts unaskedRecommend only, until a person says go.Gate · Built

L1 / The case

  1. 01ReportRecommendation and its trace.State · Built
  2. 02Handoff packetWhat a worker starts from.State · Built
  3. 03Evidence ledgerClaims, origin and limits.State · Built
Relationships, as a list
  • Report hands off to Known.
  • Cold worker hands off to Needs input.
  • Revalidation required returns to Evidence ledger.