Bob Michaels/ai
A methodology by Bob MichaelsNo. 01

HALO

Human · Agentic · Lifecycle · Orchestration

Orchestration, not autonomy.

One human drives an AI system to do the work of a team. Nothing runs unattended. The AI is the engine; I am the driver.

I · The bet

Autonomy without checkpoints burns money.

Unattended agents run past every point where a person would have said stop. Tokens spent on work nobody asked for, and bills nobody predicted.

Orchestrated agents build companies.

A human directs every step, so nothing is vibe-coded and spend follows judgment. Autonomous workers fit inside the loop: Hermes runs research and web tasks here from a written brief, and its findings pass a person before anything moves. Directing sessions or directing autonomous AI, the constant is the same. The human orchestrates the lifecycle, approvals and curation throughout.

II · The harness

Best practices in. Warm sessions out.

DISCIPLINES INSESSIONS OUT, WARMICM · structurefolders as agentarchitectureGSD · cycleplan, verify, summarize,gate: nothing silentModel routingcode first, cheap tier,frontier only for judgmentAdversarial reviewa rival model attacks thework before you see itTHE HALO HARNESSSCAFFOLDstamps every workspace the same shape:NN_stage/ · CONTEXT.md · plans/ · output/SYNCwritten at the moment of decision · commit hooksrestamp and regenerate the INDEX · status from diskTOKEN DISCIPLINEsessions read INDEX + STATE, never the corpus ·every metered call writes a ledger rowMANAGED AUTONOMYbounded workers collect from a written brief ·findings only, no side effects, a person acceptsEVERY STAGEPlanExecuteVerifySummarizeGatenothing advances until a person reads the recordWarm startoriented in the first fewreads, nothing re-derivedParallel terminalsone human, several livesessions, one protocolHuman gatesa person approves everyconsequential advanceAudit trailevery decision, spend row,and record on diskThe factory: knowledge base · templates · rules · skillsplain markdown, one git repo · learnings promote back in, so the next project starts warmer

Disciplines in · scaffold, sync, token discipline, managed autonomy · approvals throughout · learnings promote back

The model does none of the remembering. The knowledge base does none of the reasoning.

III · The cycle

Work moves in bounded stages, and a person approves every one.

Structure says where knowledge lives; the cycle says what happens next. A stage is a bounded unit with one job, one output, one evidence record, and one human gate before the work advances.

01DiscussSet the outcome and constraints.
02ResearchResolve external facts first.
03Pre-mortemName the expensive failures early.
04PlanDefine the work before execution.
05ExecuteBounded tools and agents produce the artifact.
06VerifyTest against explicit acceptance criteria.
07FixReplan on failure, never paper over it.
08SummarizePreserve what happened and what is next.
09Human gateA person approves advancement.

IV · Mission Control

One interface over every workstream.

A single cockpit pulls every project's state, files, and live terminals into one screen, and holds the queue of decisions waiting on me. Nothing consequential ships without my word, and every approval is on the record. One human, one interface, every workstream in reach.

Workstreams

4 live

Approvals

7

awaiting your word

Deploys

verified · reversible

V · The reach

The discipline is the constant. The lifecycle is the variable.

The first lifecycle

Software delivery

Git as substrateModel tieringSecurity pass every commitAdversarial reviewWatch, don't ask

No vibe coding. AI is the partner; I direct every step.

A proven second lifecycle

Editorial production

MeasureStrategizeDraftVerifyPanelEditPublishFold backthen run it again

A human in control, not just in the loop.

A proven third lifecycle

Podcast production

My essays become podcast episodes, and the whole run is orchestrated: scripting, production, mixing, art, the feed, and the cross-links between every episode and the writing it draws on. I approve every word before production and sign off on every episode before it publishes.

Script is the transcriptBroadcast-spec audioFeed on my domainEssay and episode cross-linked

The show is live. Judge it yourself. →

The same pipeline runs for clients: your existing writing becomes a podcast on a feed you own, produced on a standing cadence, with your approval at every gate. One body of work, read and heard.

VI · The disciplines

The habits that keep AI work dependable.

None are optional. Each one exists because its absence broke something.

01

Write at the moment of decision

Captured the instant it happens.

02

Read state, never re-derive it

Start warm, never cold.

03

Unknowns stay unknown

Marked, never guessed.

04

Single source of truth

One fact, one place.

05

Record everything

Feeding the system is the method.

06

Version work, lock numbers

Styling can change; figures cannot.

07

Publish nothing unverified

Every claim traces to a source.

08

Nothing autonomous unmanaged

Bounded, monitored, reviewed.

09

Adversarial review, never self-confirmation

A reviewer is asked to find failures.

10

Promote learnings only on repetition

A pattern becomes a method after it repeats.

For your organization

HALO is the method under the work, not a fourth service.

Every engagement I run, AI visibility, web transformation, or a custom build, runs on HALO. Bring in the full operating model and your team gets the same discipline, deployed on your own infrastructure.

Your knowledge base

Your company's knowledge in plain files you own, behind your firewall. The asset that appreciates, not a vendor's cloud.

Mission Control, deployed

The cockpit on your infrastructure: one interface over your projects, terminals, and approvals.

A fractional operating cadence

I run the loop with you on a standing rhythm your team can see: measure, decide, build, publish, verify, improve.

Bounded agents, human gates

Nothing autonomous runs unmanaged. Every deliverable passes a person, and every decision has an audit trail.

Training and adoption

Your people learn the session protocol and the approve-then-live loop, so you keep running it after I hand off.

Managed option

Prefer it operated for you? Keep me on to run and improve the system month to month.

Who buys the full model: teams that want the operating system behind the work, not just a one-time deliverable. Human approval, shared knowledge, bounded agents, and a full audit trail are how it reduces risk instead of adding it.

This is the short version

The manifesto has the architecture, the economics, and the receipts.