Agents¶

Agents in Arxyne operate through the three-verb model — they AUTHOR intent, they READ results, but they do not EXECUTE workflows. Execution is deterministic and agent-free.

Current Status (v0.1.0)¶

Verb	CLI	Status
READ	`--ask "..."`	Shipped. Requires an LLM backend (Ollama, Anthropic, DeepSeek, etc.)
AUTHOR	`--brief "..."`	Hidden from `--help`. Functional but not promoted — requires LLM + gateway.
EXECUTE	`--run-workflow`	Shipped. Deterministic, no LLM.

The Agent's Job¶

Phase	What the agent does	What the platform does
Before execution	Interpret an engineering brief, inspect the product, plan studies	Nothing — waits for `--run-workflow`
During execution	Nothing — not involved	Route domains, dispatch solvers, evaluate gates
After execution	Interpret results, explain failures, propose next iteration	Serve data via Python API / CLI

READ: Conversational Query¶

arxyne --ask "What's the current Cd?" --product ahmed
arxyne --ask "Which targets are failing?" --product drivaer-notchback
arxyne --ask "What happened in the last 3 runs?" --product drivaer-notchback

The agent builds a context snapshot from the USD twin (product spec, results history, competitor data, releases, provenance) and sends it to the LLM with the question. The LLM returns a natural-language answer. No tool calling — the context is pre-assembled.

--ask requires --product and an LLM backend — see LLM backends below.

AUTHOR: Engineering Brief¶

arxyne --brief "Check the drag is under target" --product ahmed

The Chief Programme Engineer (CPE) agent receives the brief and uses tool-calling to:

Run simulations (run_simulation, run_dynamics_simulation)
Assess results against gates (assess_aerodynamic_targets, assess_dynamics_targets)
Evaluate range competitiveness (get_range_competitiveness)
Generate and run workflows (generate_workflow, run_workflow)

--brief is functional but hidden from --help in v0.1.0. It will be promoted in a future release once the agent tool surface is stabilised.

Tool-Calling Architecture¶

The agent loop (arxyne_platform/core/agent.py) is a standard tool-calling loop:

Send system prompt + user message + tool schemas to the LLM
If the model returns tool calls → execute them, feed results back
Repeat until the model returns text (final answer) or max turns reached

Available tools for the AUTHOR agent:

Tool	Description
`run_simulation`	Run CFD simulation in Docker, returns Cd/Cl/CmPitch
`assess_aerodynamic_targets`	Compare aero results against USD targets
`get_range_competitiveness`	Range analysis vs benchmarked competitors
`run_dynamics_simulation`	Run Chrono dynamics (step steer, lane change, etc.)
`assess_dynamics_targets`	Compare dynamics results against USD targets
`generate_workflow`	Derive a simulation workflow from product capabilities
`run_workflow`	Execute the full multi-domain workflow

LLM Backends¶

Arxyne is LLM-agnostic. Configure via environment variables or ~/.arxyne/config.yaml:

# Local (requires Ollama + a model)
ollama pull gemma4:e4b-it-q8_0

# Cloud
export ANTHROPIC_API_KEY=***
export DEEPSEEK_API_KEY=***

The agent features (--ask, --brief) fail gracefully with an actionable message when no LLM backend is configured.

What Agents Cannot Do¶

By design, agents cannot:

Override a gate verdict (pass/fail is deterministic)
Modify the gate evaluation logic
Skip a requirement during workflow execution
Write directly to the prototype layer (only the execution harness does)

This constraint is load-bearing for certification.