How WikiSkill works
WikiSkill co-evolves agent skills with a persistent knowledge base
(wiki). Each evolution iteration runs the loop below; the wiki accumulates
across all iterations and is never rolled back.
Raw layer immutable traces ฯ ~ ฯ(x; S)
โ
Wiki Maintainer root-cause analysis โ patterns
โ
Skill Proposer ReAct over wiki + traces โ P
โ
Gating accept iff R(T_val) > R_best
Algorithm 1 (per iteration k)
Sโ = โ
, Wโ = โ
; R_best โ R_baseline(D_val, Sโ)
for k = 1..K:
if R_best == 1.0: break
T_train,k โ rollouts of ฯ with S_{k-1} # skills injected, wiki hidden
T_sample,k โ T_train,k # โค5 fails + โค3 passes (App. C)
W'_k โ M_WM(W_{k-1}, T_sample,k) # wiki maintenance (patch edits)
P_k โ M_P(W'_k, S_{k-1}, T_train,k) # ReAct proposer
S'_k โ Apply(S_{k-1}, P_k)
if R(T_val,k) > R_best: accept; S_k โ S'_k; R_best โ R(T_val,k)
else: roll back skills; S_k โ S_{k-1}
W_k โ Update(W'_k, P_k, R(T_val,k), a_k) # harness writes skill-impact.md
return S_K, W_K
Glossary
| S_k | active skill set (skills/) |
| W_k | persistent wiki (patterns, index, logs, skill-impact) |
| P_k | single atomic skill proposal (create or patch) |
| M_WM | Wiki Maintainer agent (consolidates traces into patterns) |
| M_P | Wiki-informed Skill Proposer agent (multi-turn ReAct) |
| R_best | best validation score; gating threshold (strict >) |
| R(ยท) | accuracy metric on a trace set |
Full detail in design.md, prd.md,
ui_design.md, and the paper itself.