Agent Skills

Dynamo ships agent skills: repository-native instruction files that AI coding agents such as Claude Code, Codex, and Cursor discover and follow on their own. You don’t install anything or name a skill. Clone the repo, point your agent at it, describe what you want in plain language, and the relevant skills activate.

"Optimize this Dynamo deployment for my target SLO."
"Deploy GLM-5.2 on my 4-GPU node and confirm it serves traffic."
"My time to first token (TTFT) doubled after I raised concurrency. Diagnose it and fix the config."

Deployment skills

Skills for deploying and diagnosing DynamoGraphDeployments: deploy-dynamo-recipe, troubleshoot-dynamo, dynamo-router-starter, and dynamo-interconnect-check. Ask your agent to deploy a model or investigate a broken deployment and these guide the work. Sources live under .agents/skills/.

Performance optimization skills

An end-to-end optimization loop for tuning a Dynamo deployment against a stated goal. The agent captures an immutable baseline and workload contract, deploys and smoke-tests it, benchmarks with AIPerf, proposes one evidence-backed change at a time, reviews each proposal before spending GPU time, and iterates within budgets you set (GPU-hours, wall clock, and failed deploys; the agent asks if you don’t state them). The final handoff includes the recommended configuration, reproduction commands, limitations, and preserved raw evidence.

The workflow spine is documented in agent-docs/guides/optimization/optimize-loop.md, with benchmark-validity and evidence rules under agent-docs/rules/. Harnesses that support isolated sub-agents (Codex) run the loop’s roles as separate agents; other harnesses run the same roles in one agent.

Practical notes

• For long unattended runs, the agent will tell you when to enable your harness’s goal or budget mode; if you enable it manually, wait until the agent’s opening questions are answered.
• Benchmarking and optimization consume real GPU time. State your budgets; the agent tracks consumption against them and reports it.
• Never provide secrets to the agent or allow credentials into run artifacts.