Constitutional AI self-critique against principles Building Effective Agents evaluator-optimizer pattern Claude Code Best Practices explore-plan-code workflow Visible Extended Thinking (think, think hard, ultrathink levels) Effective Harnesses one-feature-at-a-time pattern
From DeepMind:
SIMA 2 self-improvement loops Gemini Robotics hierarchical reasoning (planner + executor) Scalable AI Safety debate-based verification
From OpenAI:
Agents SDK tracing, guardrails, tripwires Deep Research adaptive planning with backtracking AGENTS.md standardized instructions
From Academic Research:
CONSENSAGENT (ACL 2025): Blind review + Devil's Advocate when unanimous. 30% false positive reduction. GoalAct: Global planning → skill decomposition → local execution. 12%+ success rate improvement. A-Mem: Zettelkasten-style memory linking for episodic→semantic consolidation. Multi-Agent Reflexion: Structured debate (Implementer → Skeptic → Advocate → Synthesizer). Iter-VF: Verify answer only, not reasoning chain. Prevents context overflow.
From Industry:
NVIDIA ToolOrchestra: Three-reward signal (outcome/efficiency/preference), dynamic agent selection AWS Bedrock: Routing mode for simple tasks, supervisor mode for complex Boris Cherny's self-verification loop (2-3x quality improvement) Simon Willison's sub-agents for context isolation
From HN discussions:
"Zero companies without human in the loop" → confidence-based escalation Context curation beats automatic RAG Fresh contexts yield better results LLM-as-judge has shared blind spots → deterministic validation
The full acknowledgements with links to every paper/resource: https://github.com/asklokesh/claudeskill-loki-mode/blob/main... Run: claude --dangerously-skip-permissions then "Loki Mode with PRD at path/to/prd" Happy to discuss any of the research or architecture decisions.
slogansand•4w ago
Effective Harnesses for Long-Running Agents - Anthropic Engineering Claude Agent SDK Overview - Anthropic Platform New Patterns:
One Feature at a Time (Rule #7 in Core Autonomy)
Work on exactly one feature per iteration Complete, commit, verify before moving to next Prevents over-commitment and ensures clean progress tracking E2E Browser Testing with Playwright MCP
Features NOT complete until verified via browser automation New Essential Pattern: Playwright MCP -> Automate browser -> Verify UI features visually Detailed verification flow added to SKILL.md Note: Playwright cannot detect browser-native alert modals Advanced Task Tool Parameters
run_in_background: Returns output_file path, output truncated to 30K chars resume: Continue interrupted agents with full context Use cases: Context limits, rate limits, multi-session work Fixed Release workflow: Use gh CLI instead of softprops action for atomic release creation