Agentic Flywheel
The Agentic Flywheel is the compounding acceleration effect that emerges when an AI agent harness reaches sufficient maturity: agents begin improving the harness itself, each improvement enables more reliable agent runs, more reliable runs generate better feedback data, and that data feeds further improvements — creating a self-reinforcing cycle.
The concept extends the Iterative-Signal-Loop from a single improvement cycle into a cumulative system. Where the signal loop describes one pass of observe → improve → validate, the flywheel describes what happens when enough passes have accumulated: momentum builds, the cost of improvement falls, and the role of humans shifts from executing improvements to governing and approving them.
The Five-Stage Model
Birgitta Boeckeler and Pete Morris (2026) define the flywheel as five progressive stages:
- Information foundation — Agents receive data to evaluate loop performance; existing tests and evaluations are the initial source
- Signal enrichment — Richer feedback is added: pipeline metrics, failure validations, production operational data
- Recommendation generation — Agents analyze results and propose improvements to upstream workflow components; the harness becomes self-evaluating
- Interactive approval — Humans-On-The-Loop review agent recommendations and direct implementation; the Why-Loop-vs-How-Loop distinction applies here — humans govern intent, agents execute
- Automated decision-making — High-confidence recommendations (scored by risk/cost/benefit) receive automatic approval; the flywheel accelerates with reduced manual overhead
Why “Flywheel” Not “Loop”
Jim Collins (2001) introduced the flywheel metaphor to describe how great companies build compounding momentum through consistent, reinforcing actions — not one-off breakthroughs. Each push of the flywheel makes the next push easier. The same physics applies to agent harness improvement:
- Each harness improvement reduces agent failure rate
- Fewer failures produce cleaner signal data
- Cleaner data enables more precise, confident recommendations
- Higher-confidence recommendations can be automated
- Automation frees human attention for higher-order governance
The result is non-linear: effort invested early compounds into capability gains that would be impossible to achieve through linear iteration.
Human Role Evolution
The flywheel fundamentally changes what humans do:
| Stage | Human Role |
|---|---|
| Early harness | Manually identify and fix failures |
| Signal loop active | Review agent runs, decide improvements |
| Flywheel spinning | Review agent recommendations, approve/reject |
| Full flywheel | Set confidence thresholds and governance policies |
This is not humans being removed — it is humans being elevated from executors to governors.
Safety Considerations
As automation of approvals increases, confidence thresholds and oversight mechanisms become critical. Weston and Foerster (2025) argue that including humans in the improvement loop provides “more transparency and steerability” than fully autonomous self-improvement. The risk of unmonitored feedback cycles is that incorrect outputs re-enter the improvement pipeline, compounding errors rather than progress. The flywheel stages are designed to prevent this: automation follows, not precedes, demonstrated human-validated confidence.
Related Concepts
- Iterative-Signal-Loop — the single-cycle foundation the flywheel builds on
- Humans-On-The-Loop — the oversight model that governs the flywheel’s automated stages
- Why-Loop-vs-How-Loop — the distinction that determines what humans approve vs. what agents execute
- Agent-Harness-Components
- Back-Pressure-Mechanisms
Sources
-
Boeckeler, Birgitta and Pete Morris (2026). “Humans and Agents.” Martin Fowler’s Bliki / Exploring Generative AI Series. martinfowler.com. Retrieved 2026-03-31. Available: https://martinfowler.com/articles/exploring-gen-ai/humans-and-agents.html
- Primary source for the five-stage flywheel model and the concept of agents recommending harness improvements
-
Boeckeler, Birgitta (2026). “Harness Engineering.” Martin Fowler’s Bliki / Exploring Generative AI Series. martinfowler.com. Retrieved 2026-03-31. Available: https://martinfowler.com/articles/exploring-gen-ai/harness-engineering.html
- “Periodic agent runs identifying inconsistencies” as flywheel component (garbage collection pattern)
-
Collins, Jim (2001). Good to Great: Why Some Companies Make the Leap… and Others Don’t. HarperBusiness. ISBN: 978-0-06-662099-2.
- Conceptual origin of the flywheel metaphor: compounding momentum through sustained, consistent reinforcing actions; contrast with the “doom loop” of companies that change direction before building momentum. Available: https://www.jimcollins.com/concepts/the-flywheel.html
-
Weston, Jason and Jakob Foerster (2025). “‘Self-Improving AI’: AI & Human Co-Improvement for Safer Co-Superintelligence.” arXiv preprint arXiv:2512.05356.
- Academic treatment of human-AI co-improvement loops; argues that including humans in improvement cycles provides “more transparency and steerability” than autonomous improvement; discusses safety risks of unmonitored self-improvement feedback cycles
-
Nakajima, Yohei (2025). “Better Ways to Build Self-Improving AI Agents.” yoheinakajima.com. Retrieved 2026-03-31. Available: https://yoheinakajima.com/better-ways-to-build-self-improving-ai-agents/
- Practitioner perspective on feedback loop design for agent self-improvement; practical patterns for structuring improvement cycles that avoid compounding errors
Note
This content was drafted with assistance from AI tools for research, organization, and initial content generation. All final content has been reviewed, fact-checked, and edited by the author to ensure accuracy and alignment with the author’s intentions and perspective.