Notes on running agents with control
Short essays on approvals, evidence, monitoring and memory: the decisions that let a team hand real work to AI agents and still account for it.
Handing work to an agent is easy. Handing it over in a way you can explain afterwards takes a few deliberate decisions: what pauses for a person, what counts as evidence, what the agent may remember and what it may spend. These notes take those decisions one at a time.
Approval and authority
- Why an agent should ask before it acts
- The seven action classes that pause for a person
- Start agents read-only
- Who may approve agent work
- Keeping money movement behind an explicit decision
Evidence and the record
- Citations that hold up
- Missing evidence is an open question
- Deciding which sources may answer
- What an agent audit trail should record
Defining the work
- Write acceptance criteria before the agent starts
- A changed question needs a new plan
- Durable task or conversation
- A refusal is a result
Watching for change
- Set a baseline before you monitor
- Review monitored changes before they travel
- From a detected change to reviewed work
Memory, tools and limits
- Memory with consent
- Bring the evidence you already have
- Connect tools without widening authority
- Governance that survives a model change
- Put a ceiling on agent spend
For step-by-step guides on research, monitoring and tool choice, see the articles. The governance page shows how these pieces fit together in GoCux.
21 resources
- Journal
A ceiling for agent spend: caps, bounds and one stop control
How to keep agent work from consuming more than you meant: visible credits, a daily cap that really stops, bounded tasks and one control that halts everything.
governance - Journal
A footnote is not a source: making agent findings checkable
How to tie each finding to the exact passage that supports it, why a link or a tidy footnote proves nothing by itself, and how reviewers should check support.
research - Journal
Approve the sources before the research starts
Why a research plan should name its allowed sources, why an empty list must mean no source at all, and what to do when the material you need is not on it.
research - Journal
Ask first, act second: the case for an approval gate
What goes wrong when agents act without a pause, what a good approval gate shows the person deciding, and why a single click is enough when the gate does its job.
governance - Journal
Decide what normal looks like before the first alert
Why a monitor needs a reviewed baseline and a written rule for material change before it runs, and why every version of that rule should be kept.
monitor - Journal
Define done first: acceptance criteria an agent can be held to
How to write acceptance criteria that someone can actually check, with weak and strong examples, and why more runtime will not rescue a vague goal for an agent.
research - Journal
Honest refusals beat spinners: error states people can act on
Why an agent that says refused or unavailable is more useful than one that spins or invents output, and how to design error states that point to a next step.
governance - Journal
Money moves need an explicit decision, not a tool call
Why researching a payment and committing money are separate jobs, why an agent's tool call never counts as authority, and what the approver needs to see.
governance - Journal
Registered is not authorized: adding tools without adding power
How to add MCP servers, skills and connectors so an agent gains only the access a task needs, and why a registered tool should never count as permission.
cli - Journal
Review a detected change before it reaches anyone else
Why a change spotted by a monitor should be reviewed before anyone shares it or acts on it, and how occurrences, materiality and notifications fit together.
monitor - Journal
Say what you could not find: keeping research gaps visible
Why research reports should keep unsupported claims and unanswered questions apart from findings, and why plausible text in a gap does more harm than a blank.
research - Journal
Start agents read-only and widen access on purpose
Why new agent access should begin with reading only, why a granted credential is not permission to write unattended, and a staged way to widen it.
governance - Journal
Start from what you know: importing evidence agents can trust
How to bring existing documents and exports into a project with origin and date attached, so agents answer from known material, and how to refresh it on purpose.
knowledge - Journal
The agent audit trail: what to record and what auditors will ask
The five kinds of event an agent audit trail must hold on one record, why sampled or rebuilt logs fail an audit, and the questions a reviewer will put to yours.
governance - Journal
The seven actions that wait for a person, one by one
A walk through the seven consequential action classes that pause agent work in GoCux: why each one stops, and exactly what the resulting decision covers.
governance - Journal
What an agent may remember, and who gets to decide
How to give agents memory without surprises: proposals versus approved memories, the narrowest scope, retention and expiry, and corrections that keep history.
governance - Journal
When a chat should become a tracked work order
How to tell when agent work has outgrown a conversation and needs a work order, with a goal, acceptance criteria, a step-by-step record and attached results.
research - Journal
When the model changes, the rules should not
Why approval, evidence and audit rules belong in the workspace rather than the model, and what may change versus what must stay fixed when you switch models.
governance - Journal
When the monitor fires: turning a change into a scoped follow-up
How to review a monitoring occurrence, then open a follow-up work order with its own goal, plan, source list and approval instead of letting a change act on its own.
monitor - Journal
When the question moves, revise the plan instead of stretching it
How scope drift creeps into agent research, why an approved plan should never quietly widen, and a practical way to revise the plan when the question moves.
research - Journal
Who gets to say yes: designing approval for agent work
How company boundaries, project scopes and named approvers fit together, why access on one surface grants nothing on another, and how to assign approval duty.
governance