Collect and analyze advances in agentic design patterns published since May 2026, compare them with the existing 29-pattern collection at zeljkoavramovic.github.io/agentic-design-patterns, suggest new candidates to add, rate each candidate 0–100 along with similarity to existing patterns, uniqueness added, and citations where available — and rate the old collection itself.
A broad web census on 2026-10-09 scanned 3,655 search-result rows across 24 query formulations and extracted 425 structured finding rows covering 2026-05-01 to 2026-10-09; these consolidated into 35 relevance-shortlisted evaluated findings, of which 24 candidates scored at least 45/100. This is a best-effort sweep of the live web, not proof that every relevant paper was indexed. Sources span many independent hosts (ACM, ACL Anthology, arXiv, practitioner reports); no dependable live citation counts were extractable for 2026 artifacts, so citation availability is shown separately — a blank means "not available", never "zero citations". Some candidate assessments are single-sourced and are judged from that one page.
X: uniqueness this candidate adds to the collection (0–100). Y: overall candidate score (0–100). Bubble size: evidence maturity. Colour: similarity to the single nearest existing pattern — blue means distinct from everything in the catalog, orange means heavy overlap with one existing pattern. Dashed line at 60 marks the "add or draft now" threshold. Hover or tap a bubble for the full assessment; all scores are editorial reviewer judgments, not measured benchmarks.
Editorial reviewer ratings of the existing 29-pattern catalog, based on the extracted inventory and the comparison corpus — judgments, not measured benchmarks. Strengths: unusually broad lifecycle coverage, a useful five-part taxonomy, and distinctive practical entries (Code-Then-Execute, Dynamic Scaffolding, Stop Hook, Spec-First Agent, Session Isolation, MCP, Resource-Aware Optimization). Weaknesses: broad umbrella entries overlap with narrower mechanisms; entries lack a consistent Context–Problem–Forces–Solution–Consequences–Known Uses–Evidence template; provenance and maturity are implicit; and transaction semantics, host-authorized tool execution, skill packaging, verifier-backed search, concurrent co-creation, turn arbitration, and explicit state dependencies are absent as first-class patterns.
Recommendation: keep the 29 stable; add a "2026 candidates" track with maturity badges; promote the seven 62+ candidates first; treat 50–59 ideas as sub-patterns or variants; add fields for nearest related pattern, distinct delta, failure modes, security implications, evidence level, date added and canonical sources; avoid adding surveys or taxonomies as operational patterns unless they supply a repeatable mechanism. Catalog source: github.com
Click a column header to sort. "Cit." marks whether live citation counts were available (none were for these 2026 artifacts). Venue text distinguishes peer-reviewed proceedings (ACM, ACL), preprints (arXiv) and practitioner posts; single-host practitioner sources back only their own row.
| Candidate | Score | Uniq | Sim | Evid | Tier | Nearest existing | Date | Cit. | Venue | Source |
|---|
The 35-row evaluation also held surveys, taxonomies and framework comparisons scoring below 45 — valuable reading, but not repeatable mechanisms. Blanks are preserved where no score was assigned.
Two discovery sweeps (2,214 + 1,441 search-result rows, 24 query formulations, overlap possible) fed semantic extraction producing 234 + 191 structured finding rows; these corpora include duplicates, surveys and noisy extractions and were used only for discovery, never counted as 425 distinct papers. Consolidation and relevance ranking yielded 35 evaluated findings; 24 scored ≥45/100 and appear above. Candidate score balances usefulness, specificity, uniqueness, evidence and generalizability; similarity is to the single nearest existing pattern (high = more overlap); uniqueness is incremental value to this specific collection; evidence rates maturity and reported validation. All are editorial assessments by one reviewer pass. Citation counts could not be reliably extracted for recent 2026 artifacts, so no candidate was ranked down for missing citations.
Data: live-web research dated 2026-10-09 covering material published 2026-05-01 to 2026-10-09; 3,655 search-result rows scanned, 425 extracted finding rows, 35 evaluated findings, 24 candidate patterns scored 0–100 (editorial judgments), compared against the 29-pattern catalog extracted from the owner's GitHub README. Scores measure editorial usefulness/uniqueness/evidence, not citations. Raw discovery corpora (234 + 191 rows) and full pattern templates were cut for space; each retained row links its source page.