Outcome-Driven Agent Loop
Give the agent a goal, tools, and context, then let it iterate to completion
- Difficulty
- Moderate
- Time to result
- ~days to results
- Steps
- 7
- Confidence
- 97%
Instead of prescribing every action in advance, define the result the agent must produce and give it the tools, context, and boundaries required to pursue that result. The agent then plans, acts, observes what happened, and iterates until the completion criteria are satisfied. This allows it to adapt when a rigid workflow would fail or when the best sequence cannot be known beforehand. Human leverage moves upstream: deciding which outcomes deserve effort, setting constraints, and judging whether progress remains useful. The operator can still nudge the loop with feedback or additional context, but avoids micromanaging each action. The model is especially useful as tool-using agents improve at constructing their own workflows from clear goals.
Origin
Extracted from Marketing Against The Grain during a discussion of loops as an alternative to explicitly programmed agent workflows.
Core principles
- 01Specify the outcome more clearly than the procedure
- 02Supply tools and context that make progress possible
- 03Let the agent adapt its process to intermediate results
- 04Use human judgment to prioritize worthwhile outcomes
- 05Nudge the loop when evidence shows it is drifting
How to run it
- 1
Choose a Worthwhile Outcome
Select a result valuable enough to justify autonomous effort and state it concretely.
Pro tip Describe the artifact, decision, or measurable state that should exist at completion.
Watch out An agent cannot compensate for poor prioritization.
- 2
Define Completion
Specify the evidence or criteria that indicate the outcome has actually been achieved.
Pro tip Use observable checks rather than subjective language such as 'make it better.'
Watch out Without a stopping condition, the loop may continue producing activity instead of value.
- 3
Supply Tools and Context
Provide the systems, data, knowledge, and permissions the agent needs to make progress.
Pro tip Include durable company context when the task spans multiple systems.
Watch out Do not grant unrelated tools or unrestricted permissions.
- 4
Set Boundaries
Identify actions that require approval, forbidden actions, budgets, and escalation conditions.
Pro tip Make boundaries explicit before the first autonomous action.
Watch out Goal pursuit without constraints can optimize the wrong trade-offs.
- 5
Run the Loop
Let the agent plan, act, inspect results, and choose the next action until the completion criteria are reached.
Pro tip Require visible intermediate artifacts or progress reports for longer loops.
Watch out Do not mistake repeated action for verified progress.
- 6
Nudge From Evidence
Review intermediate results and supply targeted corrections when the agent stalls, drifts, or uncovers new constraints.
Pro tip Change the goal only deliberately; otherwise adjust context or boundaries.
Watch out Constant intervention defeats the benefit of outcome-driven autonomy.
- 7
Verify the Result
Check the final output against the original completion criteria and retain useful learning for future loops.
Pro tip Ask the agent to summarize what worked and what should change next time.
Watch out The agent's claim of completion is not independent verification.
In the wild
A marketer asks an agent to improve a campaign toward a defined acquisition target, provides ad-platform access, brand context, a test budget, and approval boundaries, then instructs it to iterate. The agent reviews performance, proposes experiments, runs approved changes, and continues until it reaches the target or an escalation condition.
→ The agent adapts its workflow to observed campaign data instead of following a brittle predetermined sequence.
Common mistakes
Giving a Goal Without Completion Criteria
A broad aspiration does not tell the loop when to stop or how to distinguish progress from motion.
Omitting Human Prioritization
Even a capable agent needs a human to decide which outcomes are worth pursuing.
Looping Without Boundaries
Unbounded tools, spending, communications, or system changes introduce unnecessary risk.
Is it for you?
Best for
Knowledge-work outcomes that require exploration, tool use, and adaptation as new information appears.
Not ideal for
High-risk operations that require a fixed, auditable procedure or approval at every consequential step.
From the transcript
“rather than describing a sequence of steps for an agent to take, just describe a goal and give it the tools and context to nudge…”
“you give it an outcome and you tell it to like loop until you finish that outcome and it's able to like start, and you…”
“you still need to have the ideas and you have to have the prioritization of what is work worth doing.”
From the episode
Build An Army Of AI Employees To Run Your Business (With Hermes AI)