Outcome
Name the useful result before asking an agentic worker to move. What outcome would make the work worth doing?
Field guide for agentic leadership
AI changes work from doing to directing.
Directing is management. A practical field guide for leaders learning how to manage agentic workers with evidence, boundaries, cadence, attention, escalation, and human judgement.
The shift
Directing is management.
When people use agentic systems well, they are not only producing output. They are setting intent, shaping tasks, judging quality, checking evidence, protecting attention, and staying responsible for what goes out into the world.
The skill is not magic prompting. It is management as a human literacy: knowing what good looks like, knowing what you do not know, and keeping yourself meaningfully in the loop.
Field guide
The seven moves of an agentic leader.
Use these as a practical check before giving an agentic worker more autonomy, more context, or more trust.
Role
Say what the agentic worker may do, what the human owns, and who is accountable for the final decision.
Boundary
Define what it can read, change, publish, spend, remember, and escalate before autonomy gets wider.
Evidence
Ask what supports the answer, how confident you should be, what is missing, and what would change your mind.
Cadence
Create a rhythm for review, exceptions, drift, lessons, and decisions that need slower human judgement.
Attention
Protect the scarce thing. Decide where attention belongs before tools, feeds, and agents spend it for you.
Escalation
Pause when stakes rise, signals conflict, the pattern is novel, or the decision is not yours to make.
Judgement loop
Evidence is decision support, not certainty.
Use this loop before widening autonomy, accepting high-confidence output, or moving from recommendation to decision.
Head
What is the evidence, how strong is it, what is missing, and what would change our mind?
Heart
Who is affected, who has not been heard, and where does fairness matter most?
Gut
What feels off, which weak signals are present, and what could fail if we are overconfident?
Spine
Which rules, rights, thresholds, and accountabilities must hold even under pressure?
Core purpose
What human value, organisational purpose, or public good is this work ultimately meant to serve?
Decisions under scrutiny
The larger the consequence, the stronger the evidence and process should be.
Agentic work speeds up decisions. Leadership is making the uncertainty, accountability, blind spots, and review path visible before speed amplifies weak judgement.
Confidence is not certainty
Treat evidence as decision support. Ask what can be supported, with what confidence, for which decision, under which risks.
Power creates blind spots
Build routines for dissent, premortems, stakeholder challenge, and review instead of relying on senior instinct alone.
Fair process keeps respect
People may dislike a decision and still respect the process when the reasoning, constraints, and review path are clear.
Human in the loop is not enough
Oversight only works when the human has time, context, authority, evidence, and permission to intervene.
Attention and agency
The work has to contain you.
If all a person does is click the button and accept what comes back, they are not in the loop. The value is in the judgement: what they asked, what they rejected, what evidence changed their mind, and where they chose to put attention.
First steps
Start small, but manage the work properly.
This is not about banning AI or pretending it is magic. It is about showing people how to use it well enough that their own thinking becomes more visible.
Define good
Write three signs of a good answer before the agentic worker produces one.
Brief the work
Give context, audience, constraints, source expectations, and a clear stopping point.
Set the boundary
Decide what the agent can read, change, publish, remember, or escalate before the task starts.
Show the evidence
Ask for sources, assumptions, uncertainty, and rejection tests before accepting confident output.
Record judgement
Keep a small trail: what you asked, what changed your mind, what you rejected, and why.
Work harnesses
Move out of chat and into the harness.
A chat can produce words. A harness lets an agentic worker see the work, use tools, change files, run checks, create reports, and bring evidence back for review.
See the work
Local files, repos, documents, data, and context are available instead of pasted fragments.
Use real tools
Code, scripts, browsers, checks, data tools, and plugins can be used inside the workflow.
Change the artifact
The agent can update code, reports, pages, documents, and other working outputs.
Verify as you go
Diffs, tests, previews, screenshots, logs, and smoke checks make quality inspectable.
Keep the trail
Prompts, decisions, edits, evidence, and results become reviewable management records.
Starter harness links
Places to inspect the pattern.
The important question is not which logo wins. Ask what work it can see, what tools it can use, what it can change, how it verifies, and what evidence it leaves behind.
US / global
OpenAI Codex
Coding agent for real engineering work: repos, edits, reviews, parallel tasks, and shipping workflows.
Open official page US / globalAnthropic Claude Code
Agentic coding system that reads codebases, changes files, uses CLI tools, runs tests, and returns code for review.
Open official page US / globalGitHub Copilot cloud agent
Asynchronous GitHub agent with an ephemeral Actions environment for changes, tests, branches, and pull requests.
Open official page US / globalGoogle Jules
Async coding agent for GitHub repos that works in a cloud VM, produces diffs, and opens pull requests.
Open official page Europe / Czech RepublicJetBrains Junie
JetBrains coding agent for IDE and terminal workflows, with project edits, command execution, tests, and approvals.
Open official page Europe / FranceMistral Vibe
European work and code agent for long-horizon tasks, company knowledge, tools, coding, deployment, and data residency options.
Open official page Europe / Germanyn8n AI agents
Workflow harness for tool-connected agents, logic, code, integrations, human-in-the-loop guardrails, and self-hosting.
Open official page Europe / FranceDust
Multiplayer AI workspace where agents connect to company knowledge, tools, conversations, and team workflows.
Open official page Europe / SwedenLovable
Full-stack AI development platform for turning natural language into editable, deployable web applications.
Open official pageCurated starter resources
Useful places to begin in your country, organisation, school, or team.
This is a starter map, not a giant database. The point is to help leaders and managers find credible first routes into AI literacy, education guidance, and management literacy.
England
DfE: Understanding AI in education
Free support materials for school and college staff using AI safely and effectively.
Open resource ScotlandScottish AI guidance for schools
National guardrails for safe, ethical AI use while keeping teacher judgement central.
Open resource EuropeEU AI literacy and Article 4
AI Act literacy expectations for providers, deployers, staff, and affected contexts.
Open resource GlobalUNESCO AI competency frameworks
Student and teacher competency framing for education systems and curriculum teams.
Open resource OECDOECD PISA 2029 Media and AI Literacy
Assessment direction for proactive, critical engagement with digital and AI tools.
Open resource AustraliaAustralian Framework for Generative AI in Schools
Six principles for responsible and ethical generative AI in school education.
Open resource SingaporeSingapore MOE: Artificial intelligence in education
Responsible, age-appropriate AI use with agency, inclusivity, fairness, and safety.
Open resource GlobalTeachAI Guidance for Schools Toolkit
Toolkit for education leaders creating practical AI guidance and policy.
Open resourceFollow-on essay
Management is the missing literacy.
The canonical essay lives on Tonywood.org and makes the deeper argument: we should teach people how to manage themselves, tools, tasks, attention, risk, evidence, and outcomes.
Tony's writing
Four pieces behind the field guide.
Tonywood.org
Management Is The Missing Literacy
The follow-on essay: why AI makes management a general literacy, not a specialist title.
Tonywood.orgWe Need To Teach Children How To Manage Agentic Workers
The classroom argument for teaching attention, delegation, judgement, and interruption management.
Tonywood.orgWhat I Would Do If I Were at University in 2026
How students can look for human debt, learn what good looks like, and use AI without outsourcing judgement.
Tonywood.orgGetting Things Done With Agentic Workers
A practical note on roles, boundaries, success tests, escalation routes, and evidence trails.