Agent Workflows

How I would set up agents to do real leadership work in a product organization: what triggers them, what they may do alone, where a human decides, and how you know they are working.

I advise leadership teams on running AI as part of their operating model. The shift I see most teams miss: using AI is not the same as running agents. Using AI means a person pastes a prompt into a chat. Running agents means software that is triggered by events, works with defined tools and permissions, hands off to other agents, and escalates to a named human at set checkpoints. That second model needs the same management discipline as a team of people. This repo is what that discipline looks like in practice.

The agents

	Agent	Triggered when	Hands off to
Decide	Decision memo agent	A thread passes 20 messages or 7 days with no decision	Decision owner, then the decision log
	Red-team agent	A plan is marked as a one-way door	The plan's owner
Plan	Roadmap coherence agent	A planning cycle opens, or a roadmap changes	CPO
	Market landscape agent	A new product idea is logged	Discovery agent, product lead
Govern	Board pre-read agent	14 days before a board meeting	CEO or CPO
	Due diligence agent	A partnership, acquisition or board conversation is scheduled	The executive taking the meeting
Learn	Discovery agent	Interview notes are added	Product lead

How they work together

 idea logged ──> Market landscape ──> Discovery ─────────────┐
                                                             v
 thread stalls ──> Decision memo ──> owner decides ──> DECISION LOG
                                                        │     │
 one-way door ──> Red-team ──> owner decides ───────────┘     │
                                                              v
 planning cycle ──> Roadmap coherence ──> CPO          Board pre-read
                                                        (reads the log)

The decision log is the spine. Every agent that informs a decision writes to it, and the board pre-read agent reads from it. That is how you can see, over time, whether agents are making the organization decide better. (Log format: product-operating-system.)

How every agent is specified

Each agent file follows the same agent spec:

• Trigger: the event that starts it. Agents that wait to be asked are just chat.
• Inputs and tools: what it can read, and with which systems.
• Permissions: what it may do alone, what needs approval, and what it may never do.
• Hand-off and human checkpoint: who receives the output, and where a person decides.
• Escalation: when it stops and asks instead of continuing.
• Metrics: how you will know it is working.
• Where it goes wrong: its predictable failure modes, designed for in advance.

Managing agents like a team

These are the rules I recommend to leadership teams, and apply to my own work.

1. Every agent has a named human owner. Not a team. A person who answers for its output.
2. Permissions are set before it runs. Research, drafting and analysis can be autonomous. Sending, publishing, committing budget or speaking for the company never is.
3. Confidential data only goes into approved tools. Board materials, financials and customer data stay inside systems the company has cleared for them.
4. Agents are reviewed like direct reports. Quarterly, against a scorecard: is its work accepted, what does it catch, how often does it need rework, and is it making decisions faster? Agents that do not earn their place get retired.
5. Design for failure, do not discover it. Every agent here has its failure modes written down before it runs.
6. Measure decisions, not hours. Hours saved is the wrong return on AI. Decisions made better and sooner is the right one.

What does not work

• Agents that reassure. Models tend to agree with whoever set them up. Every agent here is instructed to state the case against.
• Too many tools in one agent. I once automated a daily research brief that combined live web search with structured output. It broke every other day. The simpler version held up. Every extra tool is another way to fail.
• Letting summaries replace sources. The human at the checkpoint reads the original for anything that matters.
• Agents without an owner. They drift, nobody notices, and their output quietly stops being trusted.

Related

• engine-diagnostic: an AI maturity diagnostic for leadership teams, built on my ENGINE framework
• product-operating-system: the templates, decision log and agent role card these agents work with
• More about me on my profile