GitHub - atripati/ark: AI Runtime Kernel — a context operating system for AI agents. Eliminates tool bloat, loads only what’s needed, and gives LLMs their reasoning space back.

Cut agent costs by 80–90%. Make every step verifiable. Ship agents that don't hallucinate.

ARK doesn't let the model control the system. The runtime does.

It decides which tools run, which model handles each step, how much each decision costs, and whether the output is valid — before anything reaches the user. The model's job is reduced to what it's good at: language. Everything else is governed.

┌─ ARK Agent: Task "ark-run"
│  write a function in Go that reads CSV
│
├─ Task type: coding
├─ Context: loaded 2 tools (42 tokens) [strategy: minimal]
├─ Step 1: ✓ Reasoning verified (confidence: 70%)
│  🧪 Verification: tested (score: 100%)
│  ✅ Compiled
│  ✅ Executed
│  ✅ Tests passed
│  ✅ Lint clean
├─ Step 1: COMPLETE — func readCSV(filePath string) ([][]string, error)
│
└─ Done: 1 step, 637 tokens, 5.6s | Cost: $0.002

════════════════════════════════════════════════════════════
  🧠 ARK Memory — Learning from this execution
════════════════════════════════════════════════════════════

  📥 Ingested 2 new events
  📊 Total experience: 20 memories
  🚀 Context for next run:
     Tool experience: github_search_repos — 100% success (2 uses)
     Past: 'find Python frameworks' succeeded in 2 steps, $0.005
     Past: 'write CSV reader' succeeded in 1 step, $0.002

Why ARK Exists

Current agent frameworks have a fundamental design flaw: they let the model make infrastructure decisions.

The model picks which tools to call. The model decides if the output is good enough. The model controls retry logic. This is like letting a database query decide its own execution plan.

ARK inverts this. The runtime makes every infrastructure decision. The model only does language work.

What other frameworks do	What ARK does
Dump all 140 tool schemas into prompt	Load 3 relevant tools per task (99.9% context reduction)
Use one model for every step	Route each step: cheap model for tool calls, strong for reasoning
No cost visibility until the bill	Per-decision cost graph — every step has a dollar amount
Trust model output blindly	Cognitive governor verifies every output with calibrated confidence
Every run starts from zero	Bayesian learning persists across runs — Run 2 is smarter than Run 1
Forward raw queries to APIs	Query intelligence: noise stripping, language detection, semantic scoring

What Makes ARK Different

1. Cognitive Governor

The governor is the core of ARK. It sits between every model call and the user, enforcing trust.

Task → Classify → Predict failure → Select model → Execute → Verify → Learn → Output
         ↑                                                       │
         └───────────── Registry feeds back ─────────────────────┘

Every output gets a calibrated confidence score — not a flat number, but a signal computed from model history, tool track record, response quality, and grounding:

├─ Step 1: TOOL_CALL — github_search_repos
│  ↳ ✓ Verified (confidence: 88%)     ← model proven on this tool
├─ Step 2: ✓ Reasoning verified (confidence: 87%)  ← grounded in tool data

Confidence is variable, not decorative:

85-88% → grounded reasoning with proven model+tool combo
75% → pure reasoning without tool data
50% → ungrounded (model answered without calling tools)
Below 60% → forces strong model on next step automatically

The governor also:

Predicts failures before execution — skips models with bad track records
Injects experience into prompts ("Previous attempts with this tool had failures. Be more careful.")
Tracks per-task-type performance — learns that gpt-4o-mini handles retrieval but struggles with ranking
Records task-type observations — the registry knows performance per domain, not just per model

2. Per-Step Model Routing

ARK doesn't use one model for everything. Each step gets the right model:

🧠 Model Routing:
  Step 1 [tool_call] gpt-4o-mini  (tool calls are simple, using fast model to save cost)
  Step 2 [complete]  gpt-4o       (final reasoning benefits from strong model)

  Fast model: 1 step | Strong model: 1 step

The router learns from failures. If the fast model fails on a step type, ARK promotes it to the strong model next time. Learning persists across restarts.

3. Search Intelligence (7-Phase Pipeline)

Most agent frameworks send the user's raw query to an API and hope for the best. ARK owns the entire retrieval pipeline:

"find the top 5 most popular JavaScript backend frameworks on GitHub"

Phase 1: Query Intelligence
  → Strip noise: "javascript frameworks"
  → Detect language: JavaScript
  → Add ecosystem hint: +nodejs
  → Skip API language filter for JS (TypeScript repos also needed)

Phase 2: Retrieval
  → GitHub API: sort=stars, order=desc, per_page=30

Phase 3: Language Filter
  → Accept: JavaScript + TypeScript (NestJS, Fastify are TS)
  → Reject: Java, Python, etc.

Phase 4: Junk Filter
  → Remove: awesome-lists, tutorials, cheatsheets, interview prep

Phase 5: Semantic Scoring (3-tier)
  → "web framework" in description  → 2.0× boost (Express, Fastify)
  → No framework signal             → 0.3× penalty (unknown relevance)
  → Anti-signal (ORM, CSS, testing) → 0.01× buried (Mocha, MUI)

Phase 6: Diversity Guard
  → Max 2 repos per owner (prevents Django/Django-channels clustering)

Phase 7: Simplify
  → Essential fields only → LLM explains, never selects

The LLM never decides what's relevant. The runtime ranks. The LLM explains.

4. Code Verification Engine

ARK doesn't trust generated code. It compiles, runs, and tests it before delivering.

├─ Step 1: ✓ Reasoning verified (confidence: 70%)
│  🧪 Verification: tested (score: 100%)
│  ✅ Compiled         ← go build passed
│  ✅ Executed          ← go run passed
│  ✅ Tests passed      ← auto-generated tests passed
│  ✅ Lint clean        ← go vet passed
│  ✔ code_extraction (1 code block(s) found)
│  ✔ structural_lint (0 issues)
│  ✔ constraints (0 violations)
│  ✔ compilation (compiled successfully)
│  ✔ execution (ran without error)
│  ✔ tests (auto-generated tests passed)
│  ✔ lint (0 warnings)

The verification pipeline:

Phase	What it does
Extract	Pull code blocks, auto-detect language
Auto-Fix	Fix common model errors (orphan braces, missing error handling)
Structural Lint	Check braces, parens, completeness, placeholders
Constraint Check	Over-commenting, filler comments, unused imports
Compile	`go build` / `python -m py_compile` / `node --check`
Execute	Run with 10s timeout
Auto-Test	Generate smoke tests for functions, run `go test`
Lint	`go vet` for static analysis

If code fails verification, ARK self-corrects: feeds the compiler error back to the model, forces the strong model, and retries. If it still fails after 2 attempts, ARK refuses to deliver broken code.

5. ARK Memory — Persistent Agent Experience

Every AI agent has amnesia. ARK Memory fixes it.

from ark_memory import Agent, Experience

agent = Agent("my-agent")
exp = Experience(agent)

# Agent learns from every execution automatically
exp.tool_succeeded("github_search_repos", "python frameworks", duration_ms=2500)
exp.tool_failed("web_search", "latest news", error="API key missing")
exp.strategy_learned("coding", "strip test instructions", improvement="eliminated import conflicts")

# Next run — agent queries its own experience
best = exp.best_tool_for("search repositories")
# → github_search_repos: 100% success, avg 2500ms

context = exp.execution_context("coding task")
# → Learned strategies, tool performance, failures to avoid

ARK Memory is a separate Python package (pip install ark-memory) with:

SQLite persistence — survives crashes, zero config
Semantic search via cosine similarity on embeddings
Time decay with configurable half-life
Namespace isolation (per agent, per user, per session)
Anti-redundancy deduplication
Auto-learning collector that ingests Runtime events

6. Unified Execution — One Command, Both Layers

./ark-run.sh "write a function in Go that reads CSV"

Runtime executes the task → emits events → Memory ingests them automatically. Each run makes the next one smarter.

════════════════════════════════════════════════════════════
  🧠 ARK Memory — Learning from this execution
════════════════════════════════════════════════════════════

  📥 Ingested 2 new events
  📊 Total experience: 20 memories
     Tool successes: 7
     Executions: 10

  🚀 Context for next run:
  Tool experience:
    - github_search_repos: 100% success (2 uses)
  Past execution history:
    - Task 'write CSV reader' succeeded in 1 step, $0.002
    - Task 'find Python frameworks' succeeded in 2 steps, $0.005

════════════════════════════════════════════════════════════
  Every run makes the next one smarter.
════════════════════════════════════════════════════════════

7. Decision-Level Cost Attribution

Every step has a price tag. Cost feeds back into ranking.

💰 Cost Report: ark-run
  Total Cost: $0.004840
    Input:  $0.002750 (1100 tokens)
    Output: $0.002090 (209 tokens)

  Decision Cost Graph:
    Step 1 [tool_call: github_search_repos]  $0.000990
    Step 2 [complete]                        $0.003850

5. Task Classification

ARK classifies every task before execution and adapts its behavior:

├─ Task type: ranking          ← detected from "top", "most popular"

Task Type	Behavior
ranking	Strong model for reasoning, search tool preferred
retrieval	Cheap model sufficient, list tool preferred
coding	Strong model, code-specific verification
multi_step	High effort, full verification pipeline
summarization	Medium effort, grounded check

6. Adaptive Learning

ARK remembers across runs. Tool scores evolve based on real outcomes.

RUN 1: github_list_repos = 0.55   (no history)
RUN 2: github_list_repos = 0.69   (1 success)
RUN 3: github_list_repos = 0.95   (2 successes, compounding)

RUN 1: github_search = 0.55       (no history)
RUN 2: github_search = 0.42       (1 failure, demoted)

Learning is bounded — history can't dominate. Confidence capped at 0.80. New tools get exploration bonuses. Intent-matching boosts the right tool for the right query.

By the Numbers

Metric	Raw MCP	ARK	Improvement
Context per task	60,468 tokens	~93 tokens	99.9% reduction
Cost per task	~$0.05	~$0.005	10× cheaper
Tools loaded	All 140	3 relevant	97% fewer
Steps to answer	1 (expensive)	2 (cheap + strong)	Right model per step
Verification	None	Every output	Variable confidence
Learning	None	Persistent	Run 2 > Run 1

Quick Start

git clone https://github.com/atripati/ark.git
cd ark

# No API keys needed for demos
go run ./cmd/ark bench        # see context savings (99.9% reduction)
go run ./cmd/ark demo         # see failure → adapt → recover
go run ./cmd/ark demo-learn   # see ranking improve across 3 runs

# With OpenAI (~$0.005 per task)
export OPENAI_API_KEY=sk-...
export GITHUB_TOKEN=ghp_...

# Run Runtime only
go run ./cmd/ark run agent.yaml --task "find the top 3 Python web frameworks on GitHub"

# Run Runtime + Memory together (one command)
chmod +x ark-run.sh
./ark-run.sh "find the top 3 Python web frameworks on GitHub"

# Install ARK Memory separately
cd ark-memory
pip install -e .
pytest tests/ -v   # 55+ tests

Configuration

name: my-agent
version: "0.1"

model:
  provider: openai           # openai | anthropic | ollama
  name: gpt-4o
  max_tokens: 4096
  strategy: cost_optimized   # single | cost_optimized | quality_first
  fast_model: gpt-4o-mini
  strong_model: gpt-4o

context:
  total_tokens: 200000
  strategy: adaptive
  tool_budget: 10%
  memory_budget: 10%
  conversation_budget: 35%
  max_steps: 5
  timeout_seconds: 120

memory:
  backend: file
  path: "./ark-memory.json"

Connect Any API

tools:
  - name: get_weather
    type: http
    method: GET
    uri: "https://api.openweathermap.org/data/2.5/weather?q={city}&appid=${OPENWEATHER_KEY}"
    description: "get current weather for a city"
    params:
      - city

  - name: slack_post
    type: http
    method: POST
    uri: "https://slack.com/api/chat.postMessage"
    description: "post a message to a Slack channel"
    params: [channel, text]
    headers:
      Authorization: "Bearer ${SLACK_TOKEN}"
    write: true   # requires --allow-write

ARK handles domain allowlisting, parameter validation, cost tracking, and learning for custom tools automatically.

Built-in Tools

Category	Tools	Auth
GitHub	list_repos, get_repo, list_issues, create_issue, list_pulls, get_user, search_repos	GITHUB_TOKEN (optional)
Web Search	web_search, web_search_news	BRAVE_API_KEY
File System	file_read, file_write, file_list	None
Custom HTTP	Any REST API via agent.yaml	Defined in config

12 tools across 4 categories. All ranked, learned, and cost-tracked automatically.

Safety

Safe by default. Dangerous operations require explicit opt-in.

ark run agent.yaml --task "list repos"          # ✅ reads work
ark run agent.yaml --task "create issue"        # ❌ blocked
ark run agent.yaml --task "create issue" --allow-write  # ✅ opt-in
ark run agent.yaml --task "create issue" --dry-run      # ✅ simulate

Architecture

ark/
├── cmd/ark/                    CLI entry point
│   └── main.go                 Config, agent setup, event emitter init
├── ark-memory/                 Persistent agent memory (Python)
│   ├── ark_memory/
│   │   ├── agent.py            Agent class (remember/recall/context/forget)
│   │   ├── store.py            SQLite persistence, vector search, multi-signal ranking
│   │   ├── embeddings.py       Local hash-based + optional OpenAI embeddings
│   │   ├── experience.py       Experience engine (tool tracking, strategy learning)
│   │   ├── collector.py        Auto-learning collector (ingests Runtime events)
│   │   └── types.py            Memory, RecallResult, MemoryConfig
│   └── tests/                  55+ tests
├── ark-run.sh                  Unified run script (Runtime + Memory)
├── pkg/
│   ├── config/                 YAML config + validation
│   ├── context/                Context engine (99.9% reduction)
│   │   ├── engine.go           Tool ranking (6 signals, Bayesian confidence)
│   │   └── manager.go          Context window management
│   ├── governor/               Cognitive supervisor
│   │   ├── registry.go         Model capability registry (Bayesian learning)
│   │   ├── verifier.go         Output verification (variable confidence)
│   │   ├── intelligence.go     Task classification, failure prediction
│   │   └── adapter.go          Runtime bridge
│   ├── models/                 LLM providers (Anthropic, OpenAI, Ollama)
│   ├── router/                 Per-step model routing (persistent learning)
│   ├── runtime/                Agent execution loop
│   │   ├── agent.go            Execution loop, governor integration
│   │   ├── verify.go           Code verification (compile, test, lint)
│   │   ├── quality.go          Quality layer (auto-fix, prompt optimization)
│   │   └── events.go           Event emitter (JSONL bridge to Memory)
│   ├── cost/                   Decision-level cost attribution
│   ├── store/                  Persistent learning (JSON, decay, snapshots)
│   └── tools/                  Tool implementations
│       ├── github.go           7 GitHub tools + search intelligence pipeline
│       ├── websearch.go        Brave Search (web + news)
│       ├── filesystem.go       File system (read/write/list)
│       └── custom.go           Custom HTTP tool engine

Go Runtime:  35 files | 156+ tests | Race detector clean | 12 tools
Python Memory: 9 files | 55+ tests | Zero config | 4,800 events/sec
Total: 44 files | 210+ tests | One command runs both layers

How the Scoring Works

Every tool gets a composite score from weighted signals:

Signal	Weight	What it measures
Relevance	50%	Keyword match + intent boost
Success rate	20%	Historical success/failure ratio
Confidence	5%	Data volume (capped at 0.80 to prevent bias)
Cost	-10%	Real dollar cost per call
Latency	-5%	Penalty for slow tools
Memory bonus	up to 10%	Similar query succeeded before

New tools get an exploration bonus (+0.15) so they can compete with established tools. Intent keywords ("top", "popular", "best") boost search tools by +0.40.

Production Guarantees

Guarantee	How
No hallucination when tools available	Governor blocks ungrounded responses
Code is verified before delivery	Compile → execute → test → lint pipeline
Never delivers broken code as success	Hard fail enforcement after max retries
Auto-fixes common model errors	Orphan braces, missing error handling, indentation
Variable confidence	88% grounded, 75% pure reasoning, 50% ungrounded
No invalid tool calls	RequiredParams validated before execution
No runaway loops	MaxSteps=5, TotalTimeout=120s, per-tool retry budget, max 2 self-corrections
Cost-aware	Per-decision cost graph, budget enforcement
Self-improving	Bayesian learning + experience memory, persistent across restarts
Failure prediction	Governor predicts failures before execution
Task-aware routing	Ranking tasks force strong model, tool calls use cheap model
Diversity enforcement	Max 2 repos per owner, junk filtering
Semantic scoring	3-tier relevance (boost / penalize / bury)
Experience accumulation	Every run feeds into the next via ARK Memory

Stress Tested

Sequential (20 runs):  20/20 completed, 0 crashes, 0 hallucinated data
Parallel (10 runs):    10/10 completed, 0 crashes, 0 state corruption

Failures handled correctly:
  401 (no auth)     → LLM retried with user param → succeeded
  Tool hallucinated → rejected, valid tools listed → LLM self-corrected
  Timeout           → clean termination with structured error

Roadmap

v1.0 — Cognitive Governor ✅

Cognitive governor (verifier + registry + intelligence layer)
Task classification (ranking, retrieval, coding, multi_step, reasoning)
Variable confidence (model history + tool track record + response quality)
Failure prediction (predict → avoid before execution)
Experience-aware prompting (inject failure history into prompts)
Confidence-driven routing (low confidence → force strong model)
Search intelligence (noise stripping, language detection, semantic scoring)
Diversity enforcement (max 2 per owner, junk filtering)
Context-aware learning (per-task-type performance tracking)
Intent-aware tool selection (search vs list based on query signals)
Scoring rebalance (relevance dominates, history capped)
Tool output trimming (50-70% token reduction)
156 tests, race detector clean, 12 tools

v1.1 — Code Verification + ARK Memory ✅ (current)

Code verification engine (compile, execute, test, lint)
Auto-fix: orphan braces, missing error handling, space/tab indentation
Task decomposition: strip test instructions from coding tasks
Self-correction: 2 retries with strong model, history reset
Hard fail enforcement: never deliver broken code as success
Quality layer: prompt optimization, response cleaning
ARK Memory: persistent semantic memory (Python, SQLite, zero config)
Experience engine: tool tracking, strategy learning, model performance
Auto-learning collector: ingests Runtime events automatically
Event bridge: Go Runtime emits JSONL → Python Memory ingests
Unified execution: ark-run.sh runs both layers with one command
55+ Python tests, stress tested at 4,800 events/sec
210+ total tests across Runtime + Memory

v1.2 — Adaptive Execution (next)

Multi-step adaptive chains (fail → recall experience → adapt → succeed)
Experience-aware routing (Memory informs Runtime decisions)
MCP server connector (stdio/SSE)
Auto-discover tools from MCP servers

v1.3 — Production Runtime

Streaming output
Hot-reload agent configs
Plugin system
OpenTelemetry export
go get github.com/atripati/ark library mode
DSL prototype: 10 lines replaces 1000 lines of Python+SQL+VectorDB

Contributing

ARK is designed to be the foundational runtime for AI agents.

Good first issues:

Add tiktoken-based token counting
Write MCP server connector
Add Slack tool set
SQLite store backend
Add more language detection (Rust, Swift, Kotlin)

See CONTRIBUTING.md for setup instructions.

Why "ARK"

AI Runtime Kernel.

Not a framework — frameworks give you scaffolding and hope you fill it in. Not a wrapper — wrappers add a layer and call it abstraction.

A kernel. The lowest layer that governs how intelligence is allocated, how decisions are verified, and how money is spent. Every tool call, every model selection, every output flows through ARK before it reaches the user.

The model is the CPU. ARK is the operating system.

License

Apache 2.0 — see LICENSE and NOTICE.

推荐订阅源

Hacker News: Show HN