bash # npm npm install -g loki-mode

# Homebrew brew install asklokesh/tap/loki-mode

# Docker docker pull asklokesh/loki-mode Usage:

bash # CLI wrapper loki start ./prd.md loki start --provider codex ./prd.md loki start --provider gemini ./prd.md

# Or directly in Claude Code claude --dangerously-skip-permissions # Then say: "Loki Mode" or "Loki Mode with PRD at path/to/prd.md" Human intervention controls:

touch .loki/PAUSE — Pause after the current task

echo "instructions" > .loki/HUMAN_INPUT.md — Inject guidance

Ctrl+C once to pause, twice to exit

Research foundation:

Built on 2025 patterns from OpenAI Agents SDK, Anthropic Constitutional AI, DeepMind SIMA, Chain-of-Verification (arXiv:2309.11495), and MemEvolve (arXiv:2512.18746).

Reads your PRD and breaks work into phases: Discovery, Architecture, Infrastructure, Development, QA, Deployment, and Growth

Runs the RARV cycle (Reason, Act, Reflect, Verify) continuously

Uses parallel agents for independent tasks (up to 10 Haiku agents simultaneously on Claude)

Maintains working memory across sessions via .loki/CONTINUITY.md

Includes a 7-gate quality system with 3-reviewer blind review

Multi-provider support (v5.0+):

Claude Code: Full features (parallel agents, Task tool, MCP)

OpenAI Codex CLI: Degraded mode (sequential)

Google Gemini CLI: Degraded mode (sequential)

Comments URL: https://news.ycombinator.com/item?id=46761485

Points: 1

# Comments: 2

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

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.

Comments URL: https://news.ycombinator.com/item?id=46547971

Points: 3

# Comments: 1

New version has configurable research modes. You can disable external crawling entirely and run fully offline if needed. Feedback heard.

vs other multi-agent frameworks: Most focus on chat or simple task delegation. Loki Mode runs full SDLC - from PRD to deployed product with monitoring and business ops.

vs hiring a team: Obviously humans are better for ambiguous problems. But for well-defined PRDs, this removes the "I'll get to it this weekend" bottleneck.

I reviewed outputs and approved deployments. But architecture decisions, code, tests, docs - all Loki Mode.

It's not perfect. Some rough edges. But it works and enterprises can self-host it today.

Engineering (8 types): frontend, backend, database, mobile, API, QA, perf, infra Operations (8 types): devops, SRE, security, monitoring, incident, release, cost, compliance Business (8 types): marketing, sales, finance, legal, support, HR, investor, partnerships Data (3 types): ML, data eng, analytics Product (3 types): PM, design, tech writer Growth (4 types): growth hacker, community, success, lifecycle Review (3 types): code, business, security

Agents don't step on each other. Frontend agent never thinks about database schemas. QA agent never writes deployment scripts. Domain isolation is key.

We used RARV (Retrieve, Analyze, Reason, Validate) pattern with multi-agent collaboration. Each problem gets worked by specialized agents, reviewed by 3 parallel reviewers (code, business logic, security), and only merged after consensus.

The 99.67% isn't cherry-picked. Full run against standard SWE-Bench dataset. Happy to share methodology if anyone wants to reproduce.

SWE-Bench: 99.67% (299/300 problems) HumanEval: 98.78% Pass@1 (162/164)

For context, most single-agent systems hit 30-50%. Best proprietary ones hover around 70-80%.

The difference is architecture. 37 specialized agent types across 6 swarms (engineering, ops, business, data, product, growth). Parallel 3-reviewer code review. Feedback loops that actually learn.

To stress test it, I pointed it at a blank folder and said "build a ServiceNow replacement." It ran for 19 hours and built FireLater - complete ticket management, workflows, CMDB, knowledge base, self-service portal. I wrote zero lines of code.

New in this version: - Kanban board to visualize agent actions in real-time - Perpetual improvement via self-healing feedback loops - Smarter swarm coordination

Still open source. MIT license. Still not selling anything.

Loki Mode: https://github.com/asklokesh/claudeskill-loki-mode FireLater (built by Loki Mode): https://github.com/asklokesh/FireLater

Happy to answer questions about the architecture or benchmarks.

Comments URL: https://news.ycombinator.com/item?id=46528155

Points: 2

# Comments: 5

It's a Claude Code skill that orchestrates 37 specialized AI agents (engineering, ops, marketing, legal, etc.) to take a PRD and autonomously build, deploy, and operate a product.

Key features: - Parallel 3-reviewer code review (code, business logic, security) - Multi-cloud deployment (AWS, GCP, Azure, Vercel) - State persistence for rate limit recovery - Full business ops (Stripe billing, legal docs, landing pages)

Open source, MIT licensed. Feedback welcome - still iterating.

Comments URL: https://news.ycombinator.com/item?id=46393705

Points: 4

# Comments: 8