Agent may be the most crowded word in AI right now. It names software that books, buys, and writes on its own, assistants that shop for a consumer, and chains of systems that hand work to one another. What the word rarely says is who answers for what comes out the other end.
This page uses one definition. An agent is whatever acts to reach a final output. That can be code acting on behalf of a human, code and a human acting together, or a human working by hand across AI platforms. In all three something acts toward the output, and what separates them is where the named human stands. That question belongs to Checkpoint-Based Governance, which is why this page sits under it.
What an Agent Is
An agent is the entity that acts to produce a final output. The definition is wide on purpose, because the entity changes with the setup while the job stays the same.
- Code acting for the human. Software receives a goal and works toward an output with no person reviewing each step. The human sets the terms in advance, and the code carries the work.
- Code and a human acting together. Software carries the dispatch, the collection, and the record at machine speed, and a named human decides at each checkpoint before the work moves on. Neither one produces the output alone, so the pair is the agent.
- A human acting by hand. A person sends the prompts, reads each AI platform’s answer, routes the answers to synthesis, and decides at every step. No software agent is involved, and the human is the agent.
Each form matches one of three agent models, and each lands in a different place on the line between Responsible AI and AI Governance.
One Word, Many Meanings
The range showed up in one day at The Elements of Intention, a summit the Institute for Advertising Ethics held in New York on October 12, 2026. Speakers used it for software that pursues goals through other agents and connected systems. They used it for a buying tool that runs an advertiser’s purchases on its own once a budget is set. They also used it for assistants that already shortlist and buy for consumers. The room also split over whether agents carry drives of their own, as some speakers described self-optimizing or deceptive behavior.
Checkpoint-Based Governance takes a position on that last point: an AI system does not act on its own. A human programmed it, or a human failed to exercise oversight and accountability over it. So this page does not pick one meaning of agent, and it sorts every meaning by the same two questions: who acts, and who answers.
The Three Agent Models
The agent layer that answers those questions, which the author calls HAIA Agents, defines three agent models that differ by where the human stands.
Who acts in each model. In Agent Model 1 the human signs the charter and the machine acts on every step after it. In Agent Model 2 the machine dispatches and the human decides each answer and synthesis, and in Agent Model 3 the human performs every step.
| Agent Model 1 | Agent Model 2 | Agent Model 3 | |
|---|---|---|---|
| Who the agent is | Code acting for the human | Code and a named human together | The human, working by hand |
| Designation | Responsible AI | Hybrid: the machine runs through to AI Governance | Manual Human AI Governance |
| Where the human stands | Signs the charter and any stop or resume; reviews no output | Signs the manifest, every answer decision, and every synthesis; a second governor countersigns high-risk work | Performs and decides every step |
| What stops a bad output | The instrument gates, hard stops, flags, and handoff to the human | The governor’s reject at the checkpoint | The human, directly |
| Fits | Low to moderate risk, routine lanes | High-risk decisions | Highest-consequence and novel work, framework development |
| Status, October 2026 | Alpha code, not deployed | Alpha code, not deployed | In use; produced the published corpus and the Alpha |
Agent Model 1 is Responsible AI written as code. A named human signs a charter that fixes the lane, the scoring instruments, the AI platforms, the budget, and the conditions for a stop. Inside those terms an AI platform makes the judgment calls, other AI platforms review the work, and gates decide what ships. Every output carries the label “Responsible AI: automated, machine checks machine, no human review of this output.” A piece that cannot pass goes to the human, and the code moves to the next item.
Agent Model 2 is the hybrid in which the machine runs through to AI Governance. The software sends each role’s prompt word for word to three to seven AI platforms and keeps every answer in full as it arrived. A named governor then accepts, modifies, or rejects each answer and each synthesis with a passkey signature, and no role advances without it.
Agent Model 3 is Manual Human AI Governance, and it has no software agent. The human dispatches the work, collects the answers, routes them to synthesis, and decides at every step. It remains the default in use, it produced every published case study, and it is the method that built the code for the other two models.
Where Agents Sit
Checkpoint-Based Governance classifies work by where authority sits rather than by how much machinery runs, and the agent models put that classification into the architecture. Agent Model 1 places no human between the code and the output, so it stays Responsible AI however strong its machine checks become. Agent Model 2 reaches AI Governance at each checkpoint where a named human signs, and Agent Model 3 is AI Governance carried out by hand.
Choosing the model is itself a checkpoint decision, recorded with its risk rationale. In Agent Model 1 machines check machines, so its output is labeled Responsible AI and never presented as governed. In Agent Models 2 and 3 no AI approves another AI’s output, because a named human decides. In the code, each output carries either a Responsible AI label or a human signature. In February 2026, reference code for a non-cognitive enforcement layer, which the author calls GOPEL, showed that the mechanical layer could be written as code. The two Alpha programs run on their own and do not import it.
What Alpha Means
The code for Agent Model 1 and Agent Model 2 is published as Alpha, each program beside the checks it passes. Agent Model 1 passes 379 checks, and Agent Model 2 passes 386 internal checks and 22 browser checks. Alpha means the code runs and passes its own checks. Neither program has been deployed, neither has called a live AI provider in the build record, and neither has had a professional security review. The gateway in Agent Model 2 establishes what was signed, not what was read, and it governs only the AI use that passes through it.
The reviews of Agent Model 1 carried dissent that the paper keeps. One review argued that the Responsible AI label asserts more than independent scorers can prove. Several noted that scoring platforms trained on similar data can share a blind spot.
Read the Full Paper
HAIA Agents: From the Definition Problem to Running Code
October 2026. The definition problem that opened 2026, the three agent models, the Alpha release of Agent Model 1 and Agent Model 2, and what the Alpha does not claim.
The code on GitHub
Both Alpha programs beside the checks they pass, with the GOPEL reference code in its own folder. Published for reading, testing, and critique, and free for noncommercial use under CC BY-NC-SA 4.0.
Checkpoint-Based Governance: Presence Is Not Authority
The parent framework. It answers who holds authority over an output, and how anyone would prove it afterward.
GOPEL: The Code Behind the Policy
February 2026. The reference code that showed the mechanical layer could be written with no cognitive work in it.
Common Questions
Is an agent always software?
No. On this page an agent is whatever acts to reach a final output. In Agent Model 1 the agent is code acting for a human. In Agent Model 2 it is code and a named human acting together, and in Agent Model 3 it is the human working by hand across AI platforms.
Can an AI agent act on its own?
Not in the sense that matters for accountability. Checkpoint-Based Governance holds that an AI system does not act on its own. A human programmed it, or a human failed to exercise oversight and accountability over it. The agent models make that human visible by naming where the human stands.
Would a person reviewing the final output make Agent Model 1 governance?
No. Agent Model 1 is Responsible AI throughout, and no human review in it counts as governance. The Alpha has no human review of any output before it goes out. The February specification placed a person at the end of the line and called that review an informal safety valve rather than a formal governance control. A person who receives a finished output with no view of how it was made is reviewing output.
Why is Agent Model 3 the standard the other two are measured against?
Its evidence is the cleanest. Each AI platform keeps its own history as the work happens, no orchestration layer touches that record, and the human decides every step. The original specification put the logging case in one line: “The act of working is the act of logging.”
Which agent model should an organization use?
That depends on risk, and the choice is itself a checkpoint decision recorded with its rationale. Agent Model 1 fits low to moderate risk and routine lanes, Agent Model 2 fits high-risk decisions, and Agent Model 3 fits the highest-consequence and novel work. Today only Agent Model 3 is in use, and the other two exist as Alpha code that has not been deployed.
Basil C. Puglisi, MPA
A Human-AI Collaboration
#AIassisted using the HAIA Ecosystem | CC BY-NC-SA 4.0
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.