athanor-sdk 0.8.3

Multi-agent verification orchestration -- LLMs propose, formal tools prove, every change is machine-checked

kairos

Multi-agent verification orchestration for hardware and software.

LLMs can propose RTL optimizations, but they cannot prove those optimizations are correct. A single model call that claims "area reduced by 12%" has no way to guarantee functional equivalence with the original design. Kairos closes that gap: it orchestrates specialized agents (proposer, reviewer, counterexample analyst) through formal verification tools (EBMC, yosys, Lean) in a closed loop where every proposed change is machine-checked before acceptance.

kairos optimize block.sv runs a multi-agent loop: the proposer generates candidates, EBMC k-induction proves functional equivalence, yosys measures the real area delta, and the reviewer challenges the result. What ships is not what the LLM said works -- it is what the formal tools proved correct.

Kairos is a CLI and MCP server. Connect it to Claude Code, Kiro, or Cursor, and your agent gets 40+ verification tools that return machine-checked results, not LLM opinions.

Why agents need guardrails

Every agent in the kairos loop has a different role and a different incentive. The proposer optimizes, the reviewer challenges, the counterexample analyst actively tries to break the proposal. Only changes that survive all three reach formal verification. And formal verification is not simulation or testing -- it is mathematical proof over all possible inputs.

The guardrail stack:

• Pre-verification gates reject bad proposals before they reach formal tools: structural gate (cell count must decrease), toggle power gate (no >20% power regression)
• Bounded model checking (EBMC) proves properties hold for all states up to depth k
• Equivalence checking (yosys) proves the optimized netlist matches the original
• Counterexample generation refutes invalid proposals with concrete failing traces
• Trust boundary separates LLM-touching code from certificate generation (AST-lint enforced)
• Block memory tracks refuted approaches so agents cannot re-propose known failures
• Certificate chain with sha256 provenance from source through proof to output

Focus areas

Domain	Verification engine	What kairos checks
Chip design (RTL)	EBMC, yosys	Equivalence, safety properties, area/power optimization
Accelerator kernels	NKI runtime	Custom accelerator kernel correctness
Networking protocols	EBMC, z3	Congestion control properties (BBR, CUBIC, Reno convergence)
Theorem proving	Lean 4	Proof verification, sorry closure, axiom auditing
Authorization policy	Cedar	permit/forbid policy correctness
General software	mypy, hypothesis, mutmut	Types, properties, mutation testing

Kairos is built by kairos. We use our own verification and anti-hallucination tools to develop, test, and ship kairos itself. Every bug found by kairos repair --review on our own codebase makes the tool stronger.

Quick start

pip install athanor-sdk          # imports as: kairos
kairos setup                     # zero-config wizard
kairos doctor                    # verify your environment
kairos optimize block.sv         # your first proved optimization

Commands

design -- create RTL from natural language

kairos design "8-entry FIFO with sync reset"     # spec + architecture + RTL
kairos design --spec fifo_spec.sv                 # skip spec-gen, start from SVA
kairos design "pipelined ALU" --output ./alu/     # custom output directory

optimize -- area/power reduction with formal proof

kairos optimize block.sv                          # propose + verify + measure
kairos optimize block.sv --estimate-only          # preview strategies + cost, no LLM
kairos optimize block.sv --compound 5             # 5 iterative rounds
kairos optimize block.sv --liberty cells.lib      # technology-mapped cell counts
kairos optimize block.sv --local-only             # zero network except LLM API
kairos optimize kernel.py                         # auto-detects NKI kernels

verify -- auto-detect domain, multi-layer checking

kairos verify block.sv                            # RTL: yosys + EBMC bounded model check
kairos verify src/                                # directory: walks all files
kairos verify my_cca.py                           # Python CCA: telos + types + mutation (auto-detected)
kairos verify proof.lean                          # Lean: lake build + sorry check

Auto-detects the verification domain from file content, not just extension. A Python file with networking patterns (cwnd, ssthresh, pacing_rate) gets telos protocol verification THEN Python type checking and mutation testing -- stacked, not either/or.

repair -- find + fix + verify

kairos repair block.sv                            # detect issues, propose fixes, verify
kairos repair --review block.sv                   # diagnose only, no changes
kairos repair --focus security src/               # focus on security findings

explore -- fork and compare design variants

kairos explore block.sv                           # architecture exploration
kairos explore block.sv --fork area timing power  # parallel variant comparison

doctor -- environment health check

kairos doctor                                     # check deps, license, engines, models
kairos doctor --json                              # machine-readable output

setup -- first-run configuration

kairos setup                                      # interactive wizard

How it works

1. Propose: LLM generates optimization candidates (local or cloud)
2. Verify: EBMC k-induction proves functional equivalence
3. Measure: yosys synthesis confirms real area reduction
4. Report: honest verdict with claim verification

Every measurement from a tool, not an LLM estimate. Area from yosys synthesis, equivalence from EBMC, timing from OpenSTA (when .lib provided).

Configuration (kairos.yaml)

Drop a kairos.yaml in your project root:

model: claude-sonnet-4-6

methodology:
  verification:
    require_equivalence_proof: true
    minimum_bound_k: 10
  certificate:
    require_all_properties_proved: true

agents:
  reviewer:
    enabled: true
    severity_threshold: medium

Precedence: CLI flags > kairos.yaml > environment variables > defaults.

Agent integration (MCP)

{
  "mcpServers": {
    "kairos": {
      "command": "kairos",
      "args": ["mcp", "serve"]
    }
  }
}

Works with Claude Code, Kiro, Cursor, and any MCP-compatible IDE. 41 tools available including kairos_optimize, kairos_verify, kairos_repair, kairos_explore, and kairos_doctor.

Local-only mode

kairos optimize block.sv --local-only
kairos optimize block.sv --proposer-backend local  # RTL never leaves machine

No trace upload, no telemetry. Only the LLM API call leaves your machine. Use --proposer-backend local with ollama/vLLM for full air-gapped operation.

Requirements

• Python 3.10+
• kairos setup installs everything else

Optional (for full verification):

• yosys + EBMC (hardware equivalence)
• Lean 4 (theorem proving)
• OpenSTA (timing analysis, requires Liberty .lib)

Run kairos doctor to check your setup.

Documentation

• Getting Started
• Enterprise RTL Quickstart
• Security Model

License

Commercial license required for production use. Free evaluation available. Visit athanor-ai.com for access.