Human-at-the-Edges Agent Test
Automate multistep work while keeping humans at critical control points
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 6
- Confidence
- 96%
Start by describing the outcome rather than asking an AI for a single isolated response. Reconstruct how a capable person would complete that outcome, then separate judgment-heavy decisions from mechanical research, transformation, routing, and drafting steps. An agent becomes useful when several of those steps can run without intervention while a human still initiates the workflow, supplies sensitive judgment, or approves the final result. Accuracy improves when each step has limited freedom and explicit inputs, rules, and outputs. The result is a semi-autonomous workflow that removes repetitive interaction without pretending human oversight is unnecessary.
Origin
Extracted from Marketing Against The Grain during Dharmesh Shah's explanation of when an AI agent is more appropriate than a conversational interface.
Core principles
- 01Use agents for goals that require multiple connected steps
- 02Keep humans at initiation, review, or approval boundaries
- 03Automate work only after understanding the manual process
- 04Reduce degrees of freedom to improve accuracy
- 05Prefer constrained workflows over vague one-shot requests
How to run it
- 1
Define the outcome
State the higher-order result the workflow must produce. Avoid framing the task as a single prompt or isolated AI response.
Pro tip Use an observable completion condition such as a routed CRM lead or an approved draft.
Watch out A vague goal gives every downstream step too many degrees of freedom.
- 2
Map the manual process
Write down how a person would complete the work today, including research, decisions, transformations, and system updates.
Pro tip Start with a process you already know how to perform reliably.
Watch out Do not automate a process whose successful manual version you cannot explain.
- 3
Separate judgment from mechanics
Mark which steps require human context or approval and which can run predictably without intervention.
Pro tip Look for research, classification, formatting, and routing work first.
Watch out Do not remove humans from decisions with material reputational, legal, or financial consequences.
- 4
Constrain each automated step
Give every AI action bounded inputs, explicit rules, and a defined output. Chain these narrow actions instead of relying on one broad request.
Pro tip Use deterministic business rules wherever possible and reserve the model for interpretation.
Watch out Unbounded steps compound uncertainty across the workflow.
- 5
Install human control points
Let a person initiate the run, review uncertain work, or approve the final action according to the workflow's risk.
Pro tip Record why an item was escalated so the process can improve.
Watch out Human review is ineffective if reviewers lack the evidence behind the agent's decision.
- 6
Validate before scaling
Compare the agent's results with manually completed examples and increase autonomy only after it demonstrates dependable accuracy.
Pro tip Track failures by step rather than grading only the final output.
Watch out Do not confuse a convincing demonstration with production reliability.
In the wild
An agent retrieves the newest company from HubSpot, researches it with Perplexity, applies explicit territory and industry rules, and proposes the correct owner. A salesperson reviews exceptional cases before the CRM record is updated.
→ Routine leads are researched and routed with little intervention while ambiguous cases remain under human control.
Common mistakes
Automating an unknown process
If the operator cannot describe how to complete the task manually, the agent has no stable workflow to reproduce.
Using one giant prompt
A one-shot request hides distinct decisions and makes errors difficult to isolate or correct.
Removing every human checkpoint
Full autonomy is unnecessary when initiation or final review can preserve safety at low cost.
Is it for you?
Best for
It is best for repeatable multistep work with known rules and a small number of meaningful human decisions.
Not ideal for
It is not ideal for ambiguous one-off tasks whose process and acceptable output are still unknown.
From the transcript
“the time to use it agent is when you have something where you need a human of the loop but there are lots of steps…”
“where you sort of know what to do manually and you could do it manually but you want to automate some of those steps”
“you're reducing the numbers of degrees of freedom at each step of the process”
From the episode
HubSpot Co-Founder Introduces The Future Of Ai Agents