PrerequisiteAI 202 (Agentic AI Planning & Reasoning) recommended; AI 301 (Multi-Agent Collaboration & Orchestration) recommended
0 / 5 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Explain what an Agent Harness is and how it differs from Agent logic
Describe the harness as a runtime control layer between a trigger, the agent and its tools
Distinguish Agent reasoning, Workflow logic, Harness control, Tool responsibility and Human control
Name the tool, verification, human and failure boundaries that a harness enforces
Explain what can go wrong without a harness: opaque execution, lost state, duplicated actions, cascading failures, unbounded context and runs that cannot be resumed
Choose one real system, described with simulated data, to carry through the whole course
Problem and need
8 min
The demo works. Then it runs for a week.
In AI 202 and AI 301 you designed agents that plan, reason and work in teams. A design like that can look excellent in a demo. This course asks what happens next: when the work runs for a long time, holds state, calls tools that can fail, and produces results that someone must be able to trust. The core question is: agents can already reason, plan and use tools, so why do they still need a harness?
Work scenario (simulated)
An agent processes incoming purchase requests overnight: it reads each request, checks it against the budget rules, drafts an approval note and sends a notification. On the second night a network error interrupts it after it has sent 40 notifications. Someone restarts it, and it starts again from the first request, so 40 people receive a second notification. In the morning nobody can tell which requests were finished, which were skipped, or why one request was approved twice. The agent reasoned well at every step. The problem was everything around the steps.
Nobody could see what the run had done or where it stopped
The progress of the run was not saved, so the restart began from the beginning
Repeating the run repeated actions that should happen only once
No one had defined when a person should be asked, or when the run should stop
Warning
A better prompt or a stronger model does not fix these problems. They are not about what the agent thinks. They are about how its execution is controlled. That controlling structure is the subject of this course.
Key takeaway: A capable agent is not the same as a reliable system. Reliability comes from how execution is controlled, not only from how well the agent reasons.
Core concepts
10 min
Agent logic and Agent Harness
Two different questions are being answered whenever an agent system runs. Agent logic answers: what should I do next? An Agent Harness answers: how do we safely execute, track, pause, verify, resume, recover and control what happens next? The harness is the runtime control layer that wraps the agent.
The harness layer: from trigger to tools
Trigger: a user request, a schedule, an event or an incoming document starts a run
Agent Harness: holds the State, runs the control loop, saves checkpoints, applies verification and human controls, handles retry and recovery, and records what happened
Agent / Model: decides what to do next inside the limits the harness sets
Tools / Data / Environment: the outside systems the agent acts on, reached only through the harness
Text equivalent: the diagram is a vertical stack read from top to bottom. A trigger starts a run. The run goes into the Agent Harness, which contains seven parts: State, Control Loop, Checkpoints, Verification, Human-in-the-Loop, Retry and Recovery, and Observability. The harness then calls the Agent or Model, and the agent acts on Tools, Data and the Environment. Control and results travel back up through the same harness, so nothing reaches the outside world, and nothing returns from it, without passing through the harness.
Agent logic compared with Agent Harness (simulated purchase request example)
Agent logic
Agent Harness
Main question
What should I do next?
How do we safely run, track, pause, verify, resume and recover?
Concerned with
Reasoning, planning and choosing an action
State, control, limits, evidence and human involvement
Example
Decide that this request needs a budget check before approval
Record that the request is in the budget-check step, stop after three failed attempts, and ask a person if the amount is over the limit
If it is weak
The agent makes poor choices
Good choices are still lost, repeated, hidden or left without oversight
Info
The harness is not a replacement for the agent and not a particular product. It is a design layer, and the concepts in this course apply whether you build it yourself or use a framework.
Check yourself
Which statement best describes the difference between Agent logic and an Agent Harness?
Key takeaway: Agent logic asks what to do next. The harness controls how execution is carried out, tracked and recovered, and it sits between the trigger, the agent and the tools.
12 min
Who controls what: five kinds of responsibility
When you look at a real system, each piece of behavior belongs to one of five kinds of responsibility. Being able to classify them is the first harness skill, because a control placed in the wrong layer either does not work or works only by luck.
Five kinds of responsibility (simulated purchase request examples)
Kind
What it is
Example
Agent reasoning
The model deciding, interpreting or generating
Decide that a request describes a software subscription and not a hardware purchase
Workflow logic
The designed sequence of steps for the task
First check the budget, then draft the note, then notify
Harness control
Runtime rules about execution itself
Save progress after each request; stop after three failed attempts; do not send a notification twice
Tool responsibility
What an outside system does or refuses to do
The budget system rejects a request for an unknown cost centre
Human control
A person decides, approves or overrides
A manager approves any request above the agreed amount
Belongs to the agent
Interpreting the request
Choosing the next action within the workflow
Writing the draft text
Belongs to the harness
Knowing which step the run is in
Deciding whether to retry, stop or ask a person
Recording what happened, so it can be resumed and reviewed
Workflow logic says what the steps are; harness control says how the running of those steps is governed
A Tool may enforce its own rules, but the harness must still handle what happens when the Tool says no or fails
Human control is a deliberate design element. It is not what remains when automation gives up
If a rule must hold even when the model makes a mistake, it belongs in the harness or in a tool, not in the prompt alone
Info
Some behavior can be defended in two layers. What matters is that you can say which layer owns it and what happens if that layer fails.
Check yourself
A team writes in the agent prompt: "Never send the same notification twice." Which is the most reliable design?
Key takeaway: Classify behavior as Agent reasoning, Workflow logic, Harness control, Tool responsibility or Human control. Put a rule that must always hold in the harness or a tool, not only in the prompt.
Why a harness, and its boundaries
10 min
Why a harness: what breaks without one
The harness matters because of what happens when it is missing. Compare the same simulated overnight run with and without one.
Without a harness
Execution is opaque: nobody knows what happened or where it stopped
State may be lost, so an interruption means starting over
Retries may duplicate actions that should happen once
A failure in one step can cascade into the next
Context may grow without control
Humans do not know when to step in
A long task cannot be resumed
With a harness
Each step and its outcome are recorded and can be reviewed
Progress is saved so the run can continue from a known point
Retries follow a policy and avoid repeating irreversible actions
A failure is contained and routed to recovery or a person
Working context is kept within a budget
Defined points ask a person to review or approve
A run can pause, resume or stop safely
Four boundaries a harness enforces (simulated examples)
Boundary
The question it answers
Example
Tool boundary
Which tools and actions may this run use?
The agent may read the budget table but may not change it
Verification boundary
What must be checked before a result is accepted?
A drafted note must name the cost centre and the amount before it is sent
Human boundary
When must a person review, approve or take over?
Any request above the agreed amount waits for a manager
Failure boundary
What happens when something goes wrong, and how far may it spread?
After three failed attempts the run stops and reports, without touching later requests
Info
Implementation example (not the only model): a stateful framework such as LangGraph can express a harness as a graph. There the harness state is the graph state, each step is a node, and the control that chooses the next step is an edge. This is one way to build it. The concepts in this course hold whichever tool you use, and later modules add more examples clearly labelled as examples.
Human-AI-Human: a person defines the task, the limits and the points where they will be asked, the agent works inside those limits, and a person reviews the outcome and remains responsible for it.
Check yourself
A run is interrupted halfway and the system restarts it from the beginning, repeating actions that had already taken effect. Which harness capability is most clearly missing?
Key takeaway: Without a harness, execution is opaque, state is lost and actions repeat. A harness enforces the tool, verification, human and failure boundaries, and a person stays responsible.
Practice and apply
10 min
Practice: Agent Logic or Harness?
Activity 1 gives you 8 simulated design decisions. For each, choose one: Agent reasoning, Workflow logic, Harness control, Tool responsibility or Human control, and say why. Some cases can reasonably be read two ways; what matters is that you can explain your reasoning. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.
8 design decisions (decide for yourself)
Case
Design decision
Think about
1
Decide whether an incoming email is a complaint or a question
Who or what interprets meaning?
2
Always run the budget check before the approval note is drafted
Is this the designed sequence or runtime control?
3
Save the current step after each request so a restart does not begin from the start
What must survive an interruption?
4
Reject a purchase request for a cost centre that does not exist
Which system owns this rule?
5
A manager approves any request above the agreed amount
Who holds the decision?
6
Stop the run after three failed attempts on the same step
Who limits the run itself?
7
Choose a friendly or a formal tone for the reply
Is this reasoning or control?
8
Record every tool call and its result so the run can be reviewed afterward
What makes the run understandable later?
Agent reasoning: the model interprets, decides or generates
Workflow logic: the designed order and content of the steps
Harness control: runtime rules about state, limits, checks, recovery and recording
Tool responsibility: something an outside system does or enforces
Human control: a person decides, approves or overrides
1Apply: choose one real system for the whole course Pick an agent system or agentic workflow that could run for a long time, hold state, call tools that can fail, or need review. Every later module practices on it, and it ends up in your Agent Harness & Reliability Canvas.
2Describe it with simulated data Write only what the system is for and how it works, with no real names, client personal data or company secrets. If it involves sensitive content, describe a simulated version.
3Write the goal and the first harness question One sentence for the goal, then answer: what would we most need to know, control or recover if a run of this system were interrupted halfway?
You need to pick one system for the whole course. Which is the best fit?
Key takeaway: Human-AI-Human: a person defines the task and the controls, the agent works within them, and a person reviews the result and stays responsible for it.
