Agents Are Getting Two Loops: Outcomes + Dreams
The next wave is less about “bigger context” and more about
evaluation
and
learning
baked into the workflow.
🧭 Outcomes
Define success
Give the agent a goal and a tight loop for deciding whether the goal was achieved.
Turns “do your best” into
measurable completion
Makes failures visible (not silent)
Pairs naturally with
human approval gates
😴 Dreams
Stop repeating mistakes
Synthesize past runs into reusable “what we learned” patterns.
Captures corrections as
durable rules
Compounds over time (teams get better)
Shifts effort from prompt-crafting to
teaching
The meta-trend: we’re converging on three loops:
Govern
(allow/deny + enforcement),
Evaluate
(Outcomes), and
Learn
(Dreams).
Prompted by Daniel Ballard’s “Dreams + Outcomes” framing (Claude Managed Agents)