Practice → knowledge → institution
Reflective
Bootstrapping
How agents turn practice into institutional knowledge.
The research question
What can an organization inherit from the work its agents already do?
A finished task is a deliverable. Its trajectory can also be the beginning of a better way to work.
01 · The intuition
Practice produces knowledge.
We learn by doing. That familiar relationship continues when agents do the work.
Instructions describe the world as their authors understood it. Contact with reality exposes hidden prerequisites, missing steps, local exceptions, and unstated expectations. Resolving those surprises produces new knowledge.
The final deliverable rarely preserves what was discovered along the way. That knowledge remains in the trajectory, often unused by later tasks and unseen by the person who delegated the work.
RB turns what an agent learns during work into improvements that people can understand, review, and keep.
02 · The mechanism
A learning loop grounded in real work.
Reflective Bootstrapping is a governed process in which agents turn experience from practice into lasting improvements to their behavior, including improvements to how they learn.
- 01
Practice
Act in the real world with existing knowledge.
- 02
Trajectory
Keep the actions, observations, and corrections.
- 03
Reflection
Interpret what the experience revealed.
- 04
Proposal
Identify a scoped change and its evidence.
- 05
Human review
Share the learning and decide what to adopt.
- 06
Incorporation
Make accepted knowledge available next time.
What changes?
The agent-facing operating model: instructions, procedures, quality criteria, skills, precedents, and paths to relevant guidance. The person reviewing a lesson also carries new knowledge back into the organization.
What persists?
Reviewed guidance and the reasons for adopting it. Later agents and colleagues can inherit a lesson without having taken part in the original task.
03 · Reflection
From an informative episode to a reviewable change.
Reflection lenses supply the questions. Evidence grounds the interpretation. Attribution identifies where a lesson belongs.
01
Find friction
Retries, workarounds, repeated searches, and human corrections reveal gaps between expectations and reality. Informative successes can also teach.
02
Choose reflection lenses
Ask whether the episode reveals a workflow defect, missing context, a capability limit, or an authority conflict. Diverse precedents help expand the questions.
03
Form a Knowledge Delta
State what was learned, with quoted evidence, alternative explanations, scope, and uncertainty. Keep the conditions attached to the lesson.
04
Attribute the lesson
Identify which part of the operating model should change. A missing instruction, an instruction that was never loaded, and a tool defect need different responses.
05
Prepare an improvement proposal
Show the concrete revision, its intended effect, how future work will encounter it, and what would confirm or challenge it.
06
Improve reflection itself
Rejected proposals and later outcomes expose weak lenses or attribution habits. Changes to the learning method go through the same review process.
Give each responsibility a suitable method.
Reflection concentrates on meaning. Deterministic tools verify exact evidence spans, source versions, and edit preconditions. Smaller models may assist ambiguous lookup and routine routing. A text match can locate evidence; it cannot by itself establish historical provenance or causality.
04 · Human review
Review governs learning.
It also brings knowledge back to people.
Knowledge synchronization
The agent explains what happened, what it learned, and where the lesson applies. The reviewer learns about discoveries hidden behind the deliverable and adds domain context that can correct or refine the interpretation.
Accountable adoption
People accept, narrow, defer, or reject a proposed change. A judgment deposit preserves the decision, rationale, evidence, scope, and accountable reviewer so future participants can understand and revisit it.
The result is a reviewed change to the operating model and a better-informed person. An approval click alone does not demonstrate understanding.
Permissions, audit mechanisms, and human-controlled constitutional constraints remain outside reflective modification.
05 · Put it to work
Give your agent an RB skill.
Reflect on a completed task while the experience is still available, or use a saved trajectory.
The skill guides the agent through Observe → Reflect → Attribute → Propose → Synchronize → Revisit. It produces evidence-grounded lessons and scoped changes for you to review.
Download the RB skillStart with the work you just did.
Give the skill to an agent with access to the task context, then ask:
Use Reflective Bootstrapping to reflect on this task. Identify what we learned and prepare improvement proposals for my review.
Review the evidence, interpretation, and scope before accepting persistent changes. Later work tests the lesson and can improve the reflection method itself.
A host system supplies trajectory capture and incorporation. The skill preserves uncertainty and human acceptance; empirical evaluation is pending.