docs(research): adopt the agent-genome frame + block-coordinate credit assignment#208
Merged
Conversation
…t assignment External convergence (GEPA/SkillOpt/OPRO/Reflexion/Voyager/DSPy lineage): an agent is a policy induced by editable external state; optimize the genome from trajectories. Two upgrades folded into the map: (1) the Target axis becomes the full genome decomposition (prompt · skills · tool grants · topology · memory/retrieval · routing/policy · verifier · curriculum); (2) block-coordinate credit assignment as standing discipline — attribute failures via counterfactual reruns (the /autopsy move systematized), then edit the implicated coordinate; never re-descend a flat one. Reinterprets the GEPA holdout tie as a flat COORDINATE, not a flat landscape, and publishes the measured gradient table (tool grants +70pp ≫ architecture ~20pp ≫ model > strategy > prompt ~0) as the empirical prior. Two corrections imposed on the frame before adopting its mechanisms: deployable checkers only in the reward vector; selector≠judge still binds (reflection may see the judge, steering/selection may not). Mixture-of-genomes + bandit routing noted as gated.
tangletools
approved these changes
Jun 9, 2026
tangletools
left a comment
tangletools
left a comment
Contributor
There was a problem hiding this comment.
✅ Auto-approved PR — ea61b419
Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.
tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-09T22:58:21Z
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Folds the reviewed 'trajectory-conditioned optimization of the agent genome' frame into
optimization-space.md: the Target axis upgraded to the full genome decomposition; block-coordinate credit assignment (counterfactual-rerun attribution) adopted as standing discipline; the GEPA holdout tie reinterpreted as a flat coordinate, not a flat landscape; the measured gradient table published (tool grants +70pp ≫ architecture ~20pp ≫ model > strategy > prompt ~0). Two disciplines imposed on the frame before adopting its mechanisms: deployable-checker-only reward components, and selector≠judge (reflection may see the judge; steering/selection may not). Mixture-of-genomes + bandit routing noted as gated on cheap dominance evidence.