Operon is built around the idea that agent reliability depends heavily on structure.
The library borrows from biology not as decoration, but as a source of reusable control motifs:
The recommended front door for engineers:
Shared infrastructural components such as:
NucleusMembraneMitochondriaChaperoneLysosomeIncludes:
The lower-level layer models modules, ports, wires, and topology explicitly so the library can reason about architecture instead of treating everything as an opaque chain.
v0.17 made the epistemic topology story explicitv0.18 made the practical front door thinnerv0.19 added bi-temporal memory with dual time axesv0.20 integrated bi-temporal memory into the SkillOrganism runtime as an auditable substrate, formalizing the three-layer context model (topology, ephemeral, bi-temporal)v0.21 added pattern repository, watcher component, adaptive assembly, and experience pool — the adaptive structure layerv0.22 added cognitive mode annotations (System A/B), sleep consolidation, social learning, and curiosity — the cognitive architecture layerv0.23 added developmental staging with critical periods and capability gating — the developmental layer; release integration with memory adapters and cross-subsystem tests completed the roadmapv0.24 added convergence adapters (Swarms, DeerFlow, AnimaWorks, Ralph, A-Evolve), TLA+ formal verification, co-design composition — the convergence layer (C1–C4)v0.24.1 added Scion adapter, production compilers (organism_to_swarms/deerflow/ralph/scion), distributed watcher, LangGraph node — phase C5v0.25 added evaluation harness (20 tasks, 7 configs), prompt optimization protocols, workflow generation — phases C6–C7v0.25.1 added live LLM evaluation with cross-judging across Gemini API, Claude CLI, and Codex CLI; C8 roadmapv0.26 added meta-evolution (C8 Phase A): evolves organism configurations using biological primitives at the meta-level, with dual stall detection and LLM proposerv0.27 added structural guarantee benchmarks (metabolism, quorum sensing, epiplexity) with 10M+ data points, QuorumSensingBio auto-calibration, MTORScaler, real-embedding confirmationv0.28 added categorical certificate framework (Certificate, certify(), verify()), sequential pipeline validation harness with Gemma 4 local eval, Paper 4 empirical results (error bound validated, rho=+0.751)v0.29 added DNA repair module (DNARepair: checkpoint, scan, repair, certify), categorical compiler formalization (functors with Prop 5.1), security pipeline examples, 40 HuggingFace Spacesv0.30 added end-to-end real agent evaluation (RAW vs GUARDED vs FULL), multi-model comparison (Gemma 4 + Phi-3 Mini), honest findings: state integrity 100% detection/repair, immune TP=20%/FP=0%, escalation correctly scoped to repetition detectionv0.31 added VerifierComponent (adaptive immune quality evaluation, B-cell analogy), CertificateGateComponent (pre-execution G1/S checkpoint), cross-model judging, pre-stage intervention handlingv0.32 added LangGraph compiler (organism_to_langgraph()) and DeerFlow executor — compiles SkillOrganism to those frameworks and preserves structural guarantees via functor lawsv0.33 added atomic skills catalog (seed_library_from_atomic_skills(): 5 composable patterns from Ma et al.) and harness-architecture paper updates mapping Zhou et al.’s externalization framework to the categorical Architecture triple $(G, \mathrm{Know}, \Phi)$v0.34.0–0.34.2 added parallel stage execution (concurrent stage groups via ThreadPoolExecutor with LangGraph fan-out/fan-in) plus speedup theorem validation ($\rho=1.000$ across 6 configs) and categorical verification for non-interferencev0.34.4 added Phase 2 pass/fail evaluation (SWE-bench Docker harness wrapper with per-instance eval_status) and the initial null-result write-up that v0.34.5 later reframedv0.34.5 added the patch-apply pipeline (eval/_patch_sanitizer.py: rejects malformed model output before git apply; eval/_repo_cache.py + eval/_repo_grounding.py: opt-in --grounding with shallow cloning and file-snippet injection) and the SWE-bench Phase 2 v2 grounded rerun. Paper §6.3 reframed: at 8B / Q4_K_M, diff-format discipline is the binding constraint, not file selection.The current center of gravity is pattern-first workflows with auditable state, empirically validated structural guarantees, and layered protection (state/behavioral/output).