How HAAM works with agents
AI agents expand what HAAM can examine, make, and maintain. Humans still decide what should happen.
HAAM uses AI agents as bounded collaborators across research, product work, code, editorial systems, quality review, and project discovery. Each agent has a goal, limited access, tools, evidence requirements, and an approval boundary.
The definition
An agent is a workflow with tools and boundaries, not a digital employee with unlimited authority
A useful agent can continue through several connected steps: inspect the current state, gather evidence, compare options, prepare an output, and take only the actions it has permission to take. The boundary is part of the design.
A defined goal
Every agent starts with a concrete outcome, such as finding evidence, preparing a proposal, checking a website journey, or drafting a reviewable change.
Bounded inputs
The agent receives only the project context, files, pages, data, and instructions needed for the task. More access is not automatically better access.
Tools
Depending on the job, an agent can search, inspect code, read authorized files, run checks, compare options, create drafts, or prepare a pull request.
Working memory
The agent keeps track of decisions, evidence, constraints, and unresolved questions so the work can continue as a coherent process rather than a series of isolated prompts.
An approval gate
The workflow defines what the agent may do alone, what it may only propose, and what requires an explicit human decision before anything changes.
A visible output
The result should be inspectable: sources, screenshots, findings, diffs, drafts, assumptions, confidence, and the next recommended action.
The operating loop
Frame, observe, reason, propose, approve, learn
HAAM separates understanding the task from changing the world. That creates a visible point where evidence can be checked and a human can approve, redirect, or stop the workflow.
- 01
Frame
Goal, constraints, permissions
- 02
Observe
Evidence and current state
- 03
Reason
Options, risks, missing context
- 04
Propose
A visible, reviewable output
- 05
Approve
Human decision or bounded action
- 06
Learn
Result, feedback, updated context
Where agents help
Current workflows and controlled experiments
Not every idea is presented as a finished autonomous system. HAAM distinguishes practices already used in delivery from agent products and automations that are still being developed.
In active use
Product and code delivery
Agents inspect the codebase, trace related components, propose implementation plans, write changes, and prepare reviewable commits or pull requests. Testing and production decisions remain explicit steps.
In active use
Research and synthesis
Agents gather sources, compare claims, identify contradictions, organize findings, and turn a broad question into a structured decision brief with links back to the evidence.
In active use
Editorial support
Agents help organize field notes, photographs, transcripts, source material, captions, metadata, and article structure. First-hand observation and final editorial judgment stay human.
In active use
Website quality and maintenance
Agents can look for broken links, accessibility issues, missing metadata, inconsistent copy, technical regressions, and stale pages, then turn verified findings into a prioritized fix proposal.
In active use
Project framing
Agents turn an early request into goals, assumptions, constraints, open questions, options, risks, and a preliminary scope so the first human conversation starts at a higher level.
Controlled experiment
Signal-to-project discovery
HAAM is developing agents that cluster recurring user problems, weak market signals, public complaints, and opportunity gaps, then recommend a service, product, event, article, or experiment to investigate.
Permission design
The agent gets a ladder, not a blank cheque
Permission increases only when the action is understood, authorized, observable, and recoverable. A workflow can operate at different levels for different tools.
- 01
Observe
Read authorized material, inspect public pages, run checks, and collect evidence without changing the underlying system.
Search, crawl, compare, measure, summarize
- 02
Propose
Create drafts, plans, findings, code changes, issue reports, and recommended next actions for a human to review.
Draft, prioritize, generate a diff, prepare a PR
- 03
Act within bounds
Perform low-risk, reversible actions only when the workflow has clear permission, validation, and a recovery path.
Update a branch, label an issue, refresh derived data
- 04
Stop and ask
Pause when the task becomes ambiguous, sensitive, irreversible, legally meaningful, financially consequential, or outside the agreed scope.
Publishing, outreach, payments, access, production
Human approval
Some decisions should remain deliberately slow
Speed is useful until it hides responsibility. HAAM keeps explicit human gates around actions that affect people, trust, money, rights, access, reputation, or production systems.
- Publishing material in HAAM's name
- Sending client or community outreach
- Merging and deploying material changes
- Changing prices, contracts, permissions, or payment flows
- Making legal, compliance, health, safety, or accessibility claims
- Using private or sensitive data beyond the agreed purpose
- Choosing the final creative direction or business trade-off
Design rules
The agent should make the work more inspectable, not more mysterious
HAAM treats governance, evidence, accessibility, privacy, and recovery as product-design concerns rather than paperwork added after an agent is built.
Evidence before confidence
A polished answer is not evidence. Important findings should point back to a source, test, screenshot, file, measurement, or clearly identified assumption.
Least privilege by default
An agent should receive the narrowest useful access. Authorization is designed around the task instead of giving every workflow access to everything.
Reversible before autonomous
The safer an action is to inspect and undo, the more suitable it is for automation. High-impact actions require stronger approval and recovery paths.
Uncertainty stays visible
Agents should distinguish observed facts, interpretations, estimates, and unknowns rather than smoothing them into one confident story.
Humans own consequences
AI may increase speed and coverage, but HAAM remains responsible for what is selected, communicated, shipped, and maintained.
Automation must earn its place
A workflow is useful when it reduces repetitive effort or expands what can be examined without weakening trust, craft, accessibility, or accountability.
A concrete example
The HAAM Signal Agent collects evidence before it prepares a proposal
The Signal Agent discovers important public pages, runs deterministic website checks, captures supporting evidence, and uses AI only to interpret grounded findings. A human approves the findings and outreach draft before anything is sent.
Step 1
Discover
Select the pages most connected to the user journey.
Step 2
Measure
Run accessibility, performance, technical, and journey checks.
Step 3
Interpret
Explain only findings supported by captured evidence.
Step 4
Review
Approve, edit, dismiss, or keep each proposal pending.
Frequently asked questions
Useful agents need honest limits
HAAM does not treat autonomy as the objective. The objective is better work with clear ownership, evidence, and safeguards.
Is HAAM run autonomously by AI?
No. AI agents support parts of research, design, development, editorial work, quality review, and operations. Goals, permissions, trade-offs, approvals, and responsibility remain human.
Is an agent just a chatbot?
Not necessarily. A chatbot mainly exchanges messages. An agent is a bounded workflow that can maintain task state, use tools, inspect evidence, take permitted steps, and return a reviewable result.
Does HAAM let agents publish or contact people automatically?
Not by default. External communication, public claims, publishing, and other reputation-sensitive actions require a defined approval step unless a narrow workflow has been explicitly authorized.
Can an agent replace user research or professional review?
No. Agents can broaden collection, prepare analysis, and surface patterns, but they cannot replace direct participation, contextual judgment, manual accessibility testing, expert review, or accountability.