Talk to us

Home / Orchestration

Orchestration

Agent orchestration: agent workflows with human decisions built in

We built an agent counterpart of our own lab first. It has an org chart, work queues and a review process, and it runs our operation every day. Every line is written down: one agent per station, its skills, its instructions, and the decisions it has to hand to a person. We build the same discipline for your operation.

See one lineTalk to us

Our agent counterpart, in numbers

About 200Configured agentsPlatform census · 2026-08
3 of 4Lines in productionPipeline map · 2026-09
19Stages on the development lineAs configured · 2026-09
2Revision ceilingAs configured · 2026-09
1 per lineFinal call held by a personPipeline map · 2026-09
How does a line work?

One line, start to finish

A case moves through stations. An agent that did not build the work reviews it. A person makes the final call.

Example: a community articleContent line, simplified · Sample data
  1. 1StationIntakeThe orchestrator records the brief and checks the source.Talk → article
  2. 2StationCaptureA research agent pulls what the talk says, with sources.
  3. 3StationDraftThe community writer drafts the piece from the brief.
  4. RReviewEvaluateA separate reviewer scores it and runs six checks. It never rewrites.
  5. HA person decidesPublish or returnThe editor reads the package and approves, returns or escalates.

If the review fails, the piece goes back to the writer with notes. After two failed rounds it stops looping and goes to a person.

The words we use

How a line works

A pipeline is a map of stages, and each stage has a fixed list of exits. A working stage is a station: one agent, its skills and one instruction. Someone who did not build the work holds each decision stage, and a person holds the final call on every line. A small watcher with no AI in it makes sure nobody waits on a message that was never sent.

Case
One unit of work: a content piece, a feature, a product run or an investment thesis.
Stage
A named step with a fixed list of allowed exits. The system refuses any move outside the map.
Station
A working stage bound to exactly one agent, a curated set of skills and an instruction for that stage. When a case arrives, the agent wakes up with that instruction.
Who moves the case
The agent of the stage, once it has delivered. Closing a task never moves a case on its own.
Decision gate
The agent writes a decision package and stops. An approver who did not build the work approves it or sends it back. Each line sets who holds each gate.
Human gate
A decision gate held by a person. No agent can move it.
Human input
A working stage that waits for a person's answer, such as the author's interview. Nobody approves anything there, and the agent never makes up the missing answer.
Independent review
The reviewer is a different agent from the maker. It judges and never rewrites. A check it cannot measure counts as a fail.
Bounded loops
Revision rounds stop at two. The third failure goes to a person.
Closed intake
A case with a missing input does not get in. The agent lists what is missing and stops.
Alarm clock
A plain script, with no AI, that checks every 15 minutes. When a blocked task's dependencies are done and nobody said so, it posts DELIVERED or RELEASED. Prompts forget. The script doesn't.
Evidence trail
Every station closes its task with a comment and the path to what it made. Final deliverables are attached to the case as documents.
StationReview by an agent that did not build itHuman gate
Which lines run today?

The four lines

Content, development and product build run in production. The investment line is a design.

Content · ContentOS13 stages · as configured, 2026-09

The long piece: an article, a newsletter, a chapter or a script. Publishing is a human decision, and no agent publishes.

  1. 1IntakeContent Operations Orchestrator opens the run folder and records the brief. It never produces. A source it cannot verify stops the case here.
  2. 2Capture & distillResearch Analyst records what the source says, with provenance and without interpretation.
  3. 3Author kernel & interviewEditorial Strategist picks the author's voice kernel and interviews the author by comment, including one question from the opposing view. Human input: it waits for the author's answers and never invents an opinion.
  4. 4StrategyEditorial Strategist sets the output type per channel, the audience and the thesis.
  5. 5Generate (routed)The orchestrator opens one task for the channel specialist (community, newsletter, developer forum, video, social or book) and waits.
  6. 6EvaluateEditorial Board reviewer scores each criterion from 1 to 5 and runs six independent critical checks. The minimum is 4.60. It judges and never rewrites. A person approves or returns the piece.
  7. 7ReviseThe orchestrator counts the round and routes it back to the same specialist, at most twice. A third failure goes to the decision stage.
  8. 8ReportReporting Analyst writes the run report.
  9. 9ArchiveDocumentation Curator archives the piece with its provenance.
  10. 10Decision & publishThe orchestrator prepares the package and the editor-in-chief publishes. Human gate.
  11. 11PublishedThe orchestrator records the URL, date and version.
  12. 12LearnPerformance Analyst reviews results, and only with real metrics.
  13. 13CancelledEnd state.
Development · Superpowers19 stages · as configured, 2026-09

Feature and engineering work, from admission to integration. There is one human gate on this line. Approvers who did not build the work hold the others.

  1. 1IntakeEngineering Method Owner checks each item: eight required fields, a repo that works from here, and one owning agent. It answers ADMISSION: OK or REJECTED.
  2. 2BrainstormThe Method Owner explores options with the domain owner.
  3. 3RouteThe Method Owner packages the call: architectural (needs a spec), bounded (straight to a plan) or spike (a recommendation only). An engineering-standards approver decides.
  4. 4Design & specThe Method Owner writes the design doc. Public copy goes to copy evaluation first.
  5. 5Copy evaluationQuality reviewer scores public-facing copy from 1 to 5, for at most two rounds.
  6. 6Copy revisionEditorial Strategist revises and sends it back to evaluation.
  7. 7Spec approvalAn approver who did not write the spec approves it or sends it back.
  8. 8Implementation planThe Method Owner lists work items with dependencies, tests and acceptance criteria.
  9. 9Plan approvalThe go. An approver who did not write the plan approves it or sends it back.
  10. 10Isolated workspaceThe Method Owner opens a branch and worktree, with tokens scoped to the project.
  11. 11BuildThe Method Owner splits the work into child cases, one per item, each delegated to the owning specialist. The parent moves on when the last item closes.
  12. 12Case correctionNew work items when review finds gaps. The loop is capped.
  13. 13Branch reviewQuality reviewer reviews the whole branch and approves it or returns it to build.
  14. 14VerificationQuality reviewer runs the evidence checklist. Missing evidence goes back to build.
  15. 15Product acceptanceThe product owner checks the preview against each acceptance criterion. The only human gate on this line.
  16. 16IntegrationThe Method Owner integrates. Nothing reaches main without the gate above.
  17. 17SpikeA recommendation with no code.
  18. 18DoneEnd state.
  19. 19DiscardedEnd state.

A case, station by station · Feature: export report as PDF · Sample data

  1. IntakeMethod Owner · 5 minADMISSION: OK
  2. RouteStandards approver · 4 minARCHITECTURAL, approved
  3. Spec approvalStandards approver · 5 minAPPROVE, design_spec.md v02
  4. Plan approvalStandards approver · 4 minAPPROVE, 4 work items
  5. Alarm clock11:30RELEASED: WI-03 dependencies closed 15 min ago
  6. Branch reviewQuality reviewer · round 1 of 2RETURN: missing test on an empty dataset
  7. Product acceptanceProduct owner · 20 h waitingAPPROVE: 5 of 5 criteria ticked
Product build · Product Factory10 stages · as configured, 2026-09

An agent-built product run, from activation to promotion. Without a confirmed activation the run cannot start. Its ceilings (items, minutes, revisions and tokens) come from a configuration file, never from memory.

  1. 1ActivationThe Chief of Staff agent checks eligibility, quota, target repo and risk, then asks one question. The CEO activates or confirms.
  2. 2Objective, scope & budgetThe Method Owner drafts them. A person approves or cancels.
  3. 3IntakeThe Method Owner gathers sources, cleans them and reconciles them.
  4. 4Brief approvalThe Method Owner packages the brief. A person approves or returns it.
  5. 5ContractsThe Method Owner compiles the plan contracts.
  6. 6Story graphThe Method Owner validates the graph and opens the work items as child cases.
  7. 7ExecutionThe Method Owner supervises the items, with one piece of work in progress per item.
  8. 8PromotionThe Method Owner records the rollback point first, then proposes a decision. The quality level is the lowest score. A person promotes or returns the run.
  9. 9ClosedEnd state.
  10. 10CancelledEnd state.
Investment · a design10 stages · concept, not running today

A design for investment research: theses with falsifiers, an independent contrarian review, and a committee where people decide. It recommends no asset, and execution stays off the platform.

  1. 1Thesis intakeInvestment Intelligence Analyst records the thesis, the claim it rests on, the horizon and what would prove it wrong. No falsifier, no entry.
  2. 2ResearchResearch Analyst gathers sources with provenance, a claims register and base rates. No opinion.
  3. 3Contrarian reviewA Devil's Advocate agent, separate from the researcher, writes the strongest case against.
  4. 4Risk & fitPortfolio Risk Analyst checks sizing rules, concentration, liquidity and drawdown scenarios. Undefined limits stop the case.
  5. 5Committee packageThe Chief of Staff agent assembles a three-line summary, the dossier, the contrary memo, the risk sheet and a one-line recommendation.
  6. 6Investment committeePeople approve, return for more research, or reject. No agent can move this stage.
  7. 7External executionA person executes on the external platform. The agent only records what the person reports.
  8. 8MonitoringMonitoring Agent watches the falsifier and the risk limits, and opens a review case when a trigger fires.
  9. 9Review & memoryInvestment Memory Curator records what was believed, what happened and what changed the view.
  10. 10Rejected / archivedEnd state.
Where do people decide?

Decisions

Every gate gets a package a person can decide on in minutes. Then the decision is recorded.

The decision package

A decision, packaged

What is being approved in three lines, the files, the exits and a one-line recommendation. The decision is recorded with a name, a time and a reason.

Decision required · Evaluate (Editorial Board)Sample data

1 · What is being approved

A community article drafted from a 40-minute internal talk on agent orchestration.

This is the second draft, after one revision round. The Editorial Board scored it 4.72 and all six critical checks passed.

2 · Files

  • draft_v02.md
  • evaluation_v02.json
  • knowledge_gaps.md

3 · Exits

ActionGoes toWhat happens
ApproveReportThe article moves to reporting and archiving.
Return with notesReviseThe specialist gets your notes and writes another draft.
EscalateDecisionThe round cap is reached and a person decides.

4 · Recommendation

APPROVE · confidence 0.82

Reviewed by an agent that did not write this piece.

ApproveReturn with notesEscalateYour decision is recorded with your name, the time and your reason.
You are the gate

A person makes the final call, and sometimes the call is no.

We hold our own work to the same standard. Here are three real calls from our quality gate, cleaned of internal detail. Two of them were turned back.

Turned back

Home headline draft: "AI that runs your company"

It claimed too much autonomy and implied no person in the loop. It was rewritten to say where the human decision sits.

Turned back

A product tile showed an accuracy number with no card

The number had no source, no definition of what it counted and no date. It was held until it carried all three or was labelled as sample data.

Approved

Home copy, review round 6

It cleared the editorial board with no overclaims, and every figure had a source and a date. That is the bar we hold every page to.

gate-case export v2.1, sanitised · sha256:98a2ddbe7f1321cdf80d608fcc68a7c488df3d2ec3b04e3f108b12edc952d51f · 2026-09-19

The org chart

Who works, who decides

People sit at the top and agent roles below them, grouped by directorate. Each node shows the lines where it holds a station. Reviewers never build what they review.

The agent org chart: people at the top, agent roles grouped by directorate
About 200 configured agents across our operation (internal measurement, 2026-08).
Cost per case

Cost and time, per case

Cost and cycle time for one content case, split between agent work and time spent waiting for a person. Every figure here is sample data.

26 h 53 minCycle time (sample)
13Agent runs (sample)
US$ 5.80Estimated model cost (sample)
Sample data
StageRunsAgent timeWaiting for a personEst. costSplit
Intake14 min0US$ 0.20
Capture & distill111 min0US$ 0.60
Author interview29 min6 h 04 minUS$ 0.55
Strategy16 min0US$ 0.35
Generate118 min0US$ 1.10
Evaluate, 2 rounds214 min1 h 20 minUS$ 1.30
Revise112 min0US$ 0.80
Report and archive27 min0US$ 0.50
Decision & publish13 min18 h (overnight)US$ 0.15
Learn15 min0US$ 0.25
Total131 h 29 min25 h 24 minUS$ 5.80
Agent workingWaiting for a person

Most of the elapsed time is spent waiting for a person. That is on purpose: the gates are where judgement happens.

Sample figures for illustration. Real cost per case depends on the model, the size of the task and the number of review rounds. We report it per case on request.

Put agents to work on a real problem, with your people in charge.

Bring us one process you want off your team's plate. We'll tell you if coordinated agents fit it, and what it takes to run them at a standard you can sign off on. Checking fit costs less than building, so that's where we start.

Talk to us on WhatsAppWhatsApp +61 429 107 313
or email us at contact@kaikakulabs.com