Skip to content

Prior Art & Lessons

This page anchors the external sources behind the wiki’s design decisions and maps each project’s hard-won lesson to the Ikidna concept it informs. It exists so that claims elsewhere in the wiki (“agents declare success prematurely”, “reads parallelise, writes don’t”) can be traced to evidence rather than asserted.

The field of autonomous software engineering converged on a surprising amount between 2024 and 2026. The short version: most agent capability lives in the harness and the eval loop, not the model; context is a finite resource to be managed at runtime; verification must run the artifact; coding is the hard case for multi-agent; and the durable controls are structural, not textual.


ProjectCore ideaKey lesson folded inMaps to
Cognition / Devinsingle-threaded agent, continuous contextshare full traces not messages; conflicting implicit decisions break parallel writes; <10-ACU sessions; compaction as a first-class model; review is the new bottleneckCoordination Model, Runtime Context Management
OpenHandsagent = function from event history to next eventimmutable, replayable event stream as the audit/debug substrate; delegation as an in-stream action; CodeAct (act by running code)Execution Ledger, Orchestrator
SWE-agent (Princeton)the Agent-Computer Interface is a designed artifactwindowed views, lint-on-edit, concise structured output tripled pass rates with the same modelAgent-Computer Interface
Aiderrepo map + edit formats, benchmark-drivenPageRank repo map for context selection; edit format chosen per model; “lazy coding” as a named failureAgent-Computer Interface, Knowledge Base
Cursorhybrid local/remote indexing, background agentsencrypted, no-plaintext indexing; semantic + grep beats grep-only on unfamiliar code; async isolated VMsKnowledge Base, Security
GitHub Copilot agent / Agent HQissue → PR, async, PR as the HITL checkpointthe control plane is durable, models are swappable; per-task model routing; plan-first as a discrete stageEnd-to-End Orchestration, Smart Routing
Google Julesasync cloud agent returning a PRthe plan is an amendable human checkpoint before compute is spent; standardise the output envelopePlanning and Execution
Factory.ai (Droids)delegation, not collaborationrole-bounded specialists dispatched sequentially; inner-loop to agents, outer-loop to humans; enterprise memory across toolsCoordination Model, Cohort
Claude Code / Agent SDKLLMs using tools in a loop; workflows over agentsprefer deterministic workflows with agentic steps; the 5 patterns; context engineering; sub-agents return summariesCoordination Model, Runtime Context Management
Voyagera library of learned, reusable skillsverified successes distilled into named, composable procedural memory; skills alleviate forgettingSkills
MetaGPT / ChatDev / AutoGen / CrewAI / LangGraphmulti-agent frameworkswhere naive multi-agent fails: coordination overhead, error compounding, shared-mutable-state corruptionCoordination Model

Multi-agent coordination

Harness & interface

Context & memory

Orchestration, durability & security

Autonomy, evaluation & economics


  • Benchmark numbers are disputed and methodology-dependent. Memory-system leaderboard scores in particular are non-comparable across judges and subsets (the Zep-vs-mem0 dispute is the canary). Do not select infrastructure on a leaderboard number; run a private eval.
  • Some figures are vendor-motivated (sandbox cold-start times, “X% cost reduction”, “grep burns too many tokens”). Treat them as directional.
  • A few dated claims (SWE-bench Verified deprecation timing, secret-sprawl figures) come from 2026 sources and should be re-verified before formal citation.