Prior Art & Lessons
Description
Section titled “Description”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.
Systems
Section titled “Systems”| Project | Core idea | Key lesson folded in | Maps to |
|---|---|---|---|
| Cognition / Devin | single-threaded agent, continuous context | share full traces not messages; conflicting implicit decisions break parallel writes; <10-ACU sessions; compaction as a first-class model; review is the new bottleneck | Coordination Model, Runtime Context Management |
| OpenHands | agent = function from event history to next event | immutable, 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 artifact | windowed views, lint-on-edit, concise structured output tripled pass rates with the same model | Agent-Computer Interface |
| Aider | repo map + edit formats, benchmark-driven | PageRank repo map for context selection; edit format chosen per model; “lazy coding” as a named failure | Agent-Computer Interface, Knowledge Base |
| Cursor | hybrid local/remote indexing, background agents | encrypted, no-plaintext indexing; semantic + grep beats grep-only on unfamiliar code; async isolated VMs | Knowledge Base, Security |
| GitHub Copilot agent / Agent HQ | issue → PR, async, PR as the HITL checkpoint | the control plane is durable, models are swappable; per-task model routing; plan-first as a discrete stage | End-to-End Orchestration, Smart Routing |
| Google Jules | async cloud agent returning a PR | the plan is an amendable human checkpoint before compute is spent; standardise the output envelope | Planning and Execution |
| Factory.ai (Droids) | delegation, not collaboration | role-bounded specialists dispatched sequentially; inner-loop to agents, outer-loop to humans; enterprise memory across tools | Coordination Model, Cohort |
| Claude Code / Agent SDK | LLMs using tools in a loop; workflows over agents | prefer deterministic workflows with agentic steps; the 5 patterns; context engineering; sub-agents return summaries | Coordination Model, Runtime Context Management |
| Voyager | a library of learned, reusable skills | verified successes distilled into named, composable procedural memory; skills alleviate forgetting | Skills |
| MetaGPT / ChatDev / AutoGen / CrewAI / LangGraph | multi-agent frameworks | where naive multi-agent fails: coordination overhead, error compounding, shared-mutable-state corruption | Coordination Model |
Research & Engineering Sources
Section titled “Research & Engineering Sources”Multi-agent coordination
- Cognition: Don’t Build Multi-Agents
- Anthropic: Building a Multi-Agent Research System
- LangChain: How and When to Build Multi-Agent Systems
- Why Do Multi-Agent LLM Systems Fail?: MAST (arXiv:2503.13657)
Harness & interface
- SWE-agent: Agent-Computer Interfaces (arXiv:2405.15793)
- Aider: edit formats · repo map
- OpenHands (arXiv:2407.16741)
- Anthropic: Building Effective Agents
Context & memory
- Anthropic: Effective Context Engineering
- Anthropic: Effective Harnesses for Long-Running Agents
- Claude Code doesn’t index your codebase · Why coding agents still use grep
- Recursive Language Models
- LongMemEval (arXiv:2410.10813)
- Anthropic: Agent Skills · Voyager
Orchestration, durability & security
- Diagrid: Checkpoints Are Not Durable Execution
- Temporal: Why Agentic Flows Need Distributed Systems · 11 Temporal-for-agents pitfalls
- Simon Willison: The Lethal Trifecta
- GitHub: Copilot coding agent firewall
- SANS: The Confused Deputy / Credential Broker
- Replit incident: Incident DB 1152
Autonomy, evaluation & economics
- Swarmia: Five Levels of AI Agent Autonomy
- OpenAI: SWE-bench Verified
- OpenTelemetry GenAI semantic conventions
Caveats on the Evidence
Section titled “Caveats on the Evidence”- 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.