Becoming Engineers

Barnaby Robson
September 2026

Becoming Engineers.

Part one

A new approach to knowledge work

Barnaby Robson

Part two

How do we help people make this journey?

Chau Woeste

Part three

Practical agentic engineering

Barnaby Robson

9 months, 40 billion tokens later

hover to trace · click to open · drag to spin

Terminal

barnaby@Mac ~ %

A file we
can both read.

project-notes.md

# Project notes

## The idea
Make the work easier to repeat.

## Next steps
- [ ] Try one example
- [ ] Write down what worked

Make it agent-native.

Four swaps that make a file agent-nativeWordMarkdownPlain text. Both of us read it.PowerPointHTMLA page. Both of us see it.ExcelCSV + PythonTables the code can check.FoldersYAML properties + JSONThe file says what it is.

What the agent reads well · the formats this talk and the FY27 plan were built in

File over
Folder

Project Cedar - investment thesis.md

---
deal: Project Cedar
sector: Business services
geography: Hong Kong
workstream: Commercial DD
status: In progress
owner: Barnaby
---

# Investment thesis
Revenue depends on repeat customers.

Illustrative deal note · YAML properties above Markdown

WikiLinks.

[[Customer concentration]]

- informed_by::[[Customer concentration]]

Follow the useful connection.

LOCAL GRAPH · LINKED NOTESinformed byProject CedarMarket sizeInvestment thesisProject MapleCustomerconcentrationRevenue quality

Illustrative Obsidian-style graph · fictional deal notes

“I know
Kung Fu.”

Neo opens his eyes after downloading martial arts skills in The Matrix.

The Matrix, 1999

A candy shop
for adults.

GitHub. Skills, hooks and agent workflows
from the frontier. Open source. Free.

Four adults laughing in a sweet shop, filling bags from jars under a Free Candy sign

Supplied photograph

Automation tools.

Cron job

Runs at an exact time,
again and again. It can trigger
an agent: a morning briefing.

Hook

Reacts to a trigger.
Something happens, and the
agent takes an action.

Remote agent

Always on. An agent that
lives in a server and works
while you sleep.

Cron jobs, hooks and a remote agent · the automation that outlives the session

YOLO.

--dangerously-skip-permissions

Let it get on
with the job.

Set the blast radius first.

Nike Just Do It poster showing people jumping from a cliff into water

10 billion tokens
later…

Context engineering.

Context engineering: capture, stamp, link, answerWeb clipperPDFEmailCron · 02:00Hook · on saveRemote agent · overnightCedar vaultmanagement-call.mdmarket-report.mdcedar-vdr-index.mdbuyer-email.mdsupplier-contracts.mdsite-visit-notes.mdPROMPTCustomer concentrationcedar-vdr-index.mdRevenue cut-offmanagement-call.mdKey-person dependencybuyer-email.mdMarket share erosionmarket-report.mdWorking capital buildcedar-vdr-index.mdRaw. Markdown. No properties yet.Properties stamped. Passport checked.Linked. Filed. Fresh when you open the laptop.Answered from the linked notes, with the source on every line.

Illustrative pipeline · Obsidian Web Clipper, PDF to Markdown, cron, hooks, a remote agent · fictional Project Cedar

I had nailed
context engineering.

Rocky, the rock-like alien of Project Hail Mary, arms raised against a starfield, with the word AMAZE repeated three times above

Supplied image · Rocky, Project Hail Mary

I am AI guru

Starting out“AI guru”

Illustrative confidence · learning points in story order

Supplied Big Lebowski image framed as TIME Man of the Year

403

“OAuth authentication
is currently not allowed
for this organization”

Recorded error · 5 April 2026 UTC · reconstructed display

Multi-model.

Keep the work when
the model changes.

Test the fallback before you need it.

Supplied black-and-white group photograph, used as a visual pun on multiple models

Supplied photograph · visual analogy

Margin Call API Call.

Monitoring token use and spend becomes a core competency.

Audit where
the tokens go.

Forgotten scheduled jobs took most of my allowance.

Orchestrate
and delegate.

The smart agent plans. The cheapest capable model does the work.

Avoid API
access.

A seat spends an allowance you already hold. Metered credit just goes.

Supplied still: the overnight meeting scene from the film Margin Call

Supplied still · Margin Call (2011) · the bill I could not explain: image generation routed to a metered API.

68 of 127

skill descriptions
were being dropped.

Skill-listing diagnostic · 12 May 2026

A cathedral.
In one prompt.

Sagrada Família surrounded by construction cranes

The cathedral plan, as a visual metaphor.

Agree what good looks like.

The brief

What are we trying
to achieve?

One slice

Show me one complete
example.

The check

Does this do the job
we agreed?

The approved Neo slide from the first slice of this talk

We started this talk
with a slice.

The approved first slice of this presentation · September 2026

Deterministic v stochastic.

Scripts set the bounds. The model chooses what goes inside.

A mock slide: fixed geometry contains a model-selected chart and commentaryDeterministic · fixed by scriptsStochastic · chosen by the modelQuarterly reviewLOGOCommentaryselected bythe model.Font sizeLogo locationText and chart positionsFixed character limitChart typeCommentaryThe model writes insidethe fixed text box.

Did the change improve the work?

A/B test · the same job, with and without the change

Priority-tweets jobBaselineRound 1Round 2
Token budgetPassPassPass
QualityFailFailPass
PresentationFailPassPass

Diff · read what changed before you accept it.

Recorded comparison · 25 June 2026 · pass/fail against my own rubric

The review seat sits with you.

2,797

Source total

2.80

Rendered total

Then open the file that runs the agent.
AGENTS.md. Once a week.

Recorded totals · 9 June 2026

Blast radius: how far a mistake travels.

Once an agent can act, control becomes the hard problem. Before granting a new power, know how far its worst mistake could travel. Then reduce its scope.

Concentric scopes of an agent action, two reported incidents and four controlsAgentactionNamed scopeConnected systemsAWS / KiroA coding agent reportedly deleted a live AWS productionenvironment, causing a 13-hour outage.Cursor / PocketOSAn agent wiped a company database and its backupsin nine seconds.AllowlistsTouch only whatis named.Confirmation gatesDestructive stepswait for a yes.Narrow scopesOne job, one folder,one credential.Crawl–walk–runAdd powers in stages.Watch each stage.

PC Gamer on AWS/Kiro; Amazon attributes the outage to misconfigured access controls. Tom’s Hardware on Cursor/PocketOS.

Security.

The agent guardrails say no to five things.

Wipe the disk.
Delete the home folder.
Force over shared history.
Make the private repo public.
Dump the passwords.

A film still of a marshal standing in a doorway, arms folded, blocking the way

brun-guard, 14 September 2026 · after David Ondrej’s agent guardrails · still supplied

“Is this good?”

“Excellent.
Ready to share.”

The taste seat stays with me.

Supplied wardrobe photograph: clothes on a rail, chosen by eye

Illustrative exchange · the judgement lesson comes from the May essay · image supplied

Show me.

See it, then challenge it.
A format we can both read.
Increasingly, the format.

Click an option. Then press D. A slide cannot do this.

cedar-spa-options.html

Earn-out: three ways
to bridge the price gap

Project Cedar · SPA negotiation · click an option

Illustrative HTML view · fictional deal · Project Cedar is invented

Neo surrounded by Agent Smith copies in The Matrix Reloaded

Cognitive overload.

The Matrix Reloaded, 2003

I had built
myself
a second job.

A sysadmin. For myself. Unpaid.

I am AI idiot.

Illustrative portrait supplied by Barnaby: serving a takeaway bag at McDonald’s

40 billion
tokens later…

Voice.

You can talk 3x faster
than you type.

Speak, don’t type.

Supplied still: a man talking to his agent from bed, from the film Her

Supplied still · Her (2013), as a visual analogy

Rewire your brain.

How do we always deploy new technologies?

Absorb

Doing the old things better with the new technology

Innovate?

Things that are only possible with the new technology

Disrupt?

Redefine the question

Frame and wording after Benedict Evans, “What’s this for?”, June 2026.

Systems thinking.

The Apollo 17 Blue Marble: one connected wholeThe skill underneath it all.A system is the pattern of how its partsaffect each other over time. Change oneand the others react. Miss a connectionand the pattern breaks.Senior engineers built this senseover years of failure.With agents, you need it on day one.NASA, Apollo 17, AS17-148-22727, 7 December 1972.

After the framing in Machinations,
“Is this the only skill left?” (April 2026).

Three questions, before anything runs.

Systems thinking made concrete. Answer each question from the architecture before a step runs.

State, Feedback and Deletion: questions to answer from the architecture01 · StateWhere does state live?Who owns the truth?If two parts each believe they ownthe truth, a bug already exists.It will appear when that conflictis triggered.02 · FeedbackWhere does feedback live?How do you know it’s working?Logs, metrics and errors mustreach somewhere you look.A silent system can hide failure.03 · DeletionWhat breaks if I delete this?Can you trace it in your head?Pick a component. Trace how farits removal affects the systembefore touching it. “I don’t know”is the homework list.

The three questions follow Hak (Machinations), “Is this the only skill left?” (April 2026).

Here it is, running.

bvest. My own research system. Nine months of it.

bvest end to end: capture, organise, process, execute, and the loop backCAPTUREORGANISEPROCESSEXECUTE→→→Fiscal.ai / Capital IQ · fundamentalsFMP · pricesFRED · macroPolymarket · oddsX · postslast30days · Reddit, HN, YouTubebvest vaultequitiescryptocommodscashsignal ledgerwatchliststrategistanalystsenior analystquantmacro strategistICPMMorning brief · 06:30, TelegramTear sheet · FigmaIC memoThe bookCron, every morning. No model involved.Every file stamped and linked. One ledger.Seven seats. Each one a skill.On my phone before the market opens.Every decision re-enters as information.VISTan oil producerA signal in the feed.Filed under equities, linked to the oil thesis.Screened, deep-dived on the fiscal.ai databook, base-rated, voted, sized.In the book, on the tear sheet, in the brief.

bvest · the author’s own research system · sources as named · one holding shown, no position data

Building in teams.

Everyone pulls from the source of truth. Anyone can suggest a change by branch.

Everyone pulls from a shared repository; anyone can propose a branch for review and mergeEveryoneTeam memberTeam memberTeam memberShared GitHub repositorySource of truthmainPULLEveryone starts here.BranchSuggest a changeReviewCheck the changeANYONEMergeafter review

When ideas
have sex.

Matt Ridley. His name for how innovation actually happens.

Do not productise.
A black box cannot be forked.
Your edge is the rate the file improves.

Matt Ridley on a TED stage, a projected network image behind him

Matt Ridley, “When ideas have sex”, TED 2010, and The Rational Optimist (2010) · photograph supplied

Still learning.

My confidence: rise, collapse and a modest recovery, with the lesson at each stepHow much I thought I knewCommand lineAgent-native formatsEvery file's passportConnected knowledgeSkillsAutomationYOLOCapture, stamp, linkContext engineeringAI guruTest the fallbackToken efficiencyPrune the contextThin slicesDeterministic v stochasticA/B and evalsThe review seatBlast radiusSecurityHuman judgementShow meCognitive overloadAI idiotSpeak, don’t typeRewire your brainSystems thinkingBuilding in teamsStill learningAnyone whose work passes through a digital interface, which means everyone, will have to knowwhat a skill is, what a hook does, what context engineering means, and why blast radius matters.We are all going to have to become engineers.Confidence is illustrative; spacing follows the story. Grey labels were learnt the hard way.

You wake up. Friday this time. The server is quiet.

I am one person doing this
and finding it hard.

The corporations that have not started are running out of time to find it hard cheaply.

We are all going to have to become engineers · barnabyrobson.org, May 2026

How do we help people
make this journey?

Chau Woeste

Partner, People in M&A · KPMG UK

The gap inside one organisation

Access to the same tools can produce different work practices.Illustrative distribution70%little or no use20%occasional use5–10%advancedfeaturesApproximate use reported by one practitionerSeparate reported rollout10%of users90%of the tokensToken use measures activity andcost. It gives no measure ofbusiness value.So what do we do about it?

Approximate use split and the token example: Vas (@vasuman), “AI Adoption Is a Myth”, August 2026 (Core clipping). Practitioner account of deployments; no representative survey. Distribution shape is illustrative with no underlying series.

Demand for AI skills shifted towards senior roles

US and UK job postings, January 2024 to January 2026.Executives: 572Managers: 171Entry level: 88January 2024index = 100January 2026Endpoint comparisonAI-skill postings+57%Other postings+2.4%Fewer entry-level opportunities may reduce the future supply of experienced staff.

AI’s workforce collision, kpmg.com (KPMG / Revelio, US and UK job postings, January 2024 to January 2026). Analyses by Ailsa Maclean and Ankush Sharma. Endpoints shown; the monthly series is in the article.

AI use varies with age

These datasets cover different populations and periods.Job postings by role: US and UK, January 2024 to January 2026-12%Entry level+71%Managers+472%ExecutivesAI use by age: 11 euro-area countries, May 202447%Ages 18–3430%Ages 35–5423%Ages 55–74Assess the skills of the people who will make decisions about AI.

Left: KPMG / Revelio job postings by role, US and UK, January 2024 to January 2026. Right: ECB Blog, 21 March 2025, survey of workers in 11 euro-area countries, May 2024. ecb.europa.eu/press/blog/date/2025/html/ecb.blog20250321~6af1337b6b.en.html

Build or buy?

Firms can recruit experienced staff and develop theirexisting teams.The decision depends on the skills required and the work available through whichpeople can learn.Which capabilities should we recruit, and which should we develop through work?

AI’s workforce collision, kpmg.com. Analyses by Ailsa Maclean and Ankush Sharma.

Different roles need different support

UK indicators for roles that require AI skills.Market tightness is a proxy for hiring difficulty. A higher score indicates greater difficulty.Entry-level staff28%market tightness30days to fillGive staff work that developstheir skills. Provide anexperienced reviewer.Managers78%market tightness51days to fillGive managers time to testworkflows and help their teamsuse them.Leaders90%market tightness63days to fillHelp leaders assess where AI canchange work and who remainsaccountable.Recruitment and staff development can support the same plan.

KPMG / Revelio UK market tightness and time to fill, for roles requiring AI skills. Market tightness is a proxy for hiring difficulty.

Red teaming

Create a separate team to test an AI-based version of the business.Define the taskAsk the team to develop abusiness model that couldcompete with the currentbusiness.Give it room to testGive the team enoughindependence to test methodsthat differ from currentpractice.Plan how to use the resultsDecide who will assess themethods and help the operatingbusiness apply them.Assess what changes in the operating business.

A proposed approach discussed with Barnaby Robson. Separate-unit logic after Clayton Christensen, The Innovator’s Dilemma.

What we look for now

Use a piece of work to assess how someone makes and explains decisions.Systems thinkingExplain the full workflow andthe decisions within it.Identify where a changeaffects other people orsystems.Judgement and tasteExplain the standard used toassess an output. Identifyerrors, omissions and reasonsto revise it.CommunicationExplain the recommendation andits limits to the person whomust act on it.Look for people who can direct the work, check it, and explain it.

Assessment approach proposed for discussion.

How junior training may evolve

As firms automate some tasks in knowledge work, junior rolesmay change. Juniors may spend more time directing and checkingAI-assisted work.As AI takes on more junior work, how willpeople gain the experience to judge itsoutput?

Proposed direction for discussion.

← → to navigate
Menu