GMAN · agent operations · public view
An AI agent that reads the industry’s news, decides what’s worth publishing, and asks a human before it posts.
runs unattended on GitHub Actions, every day · recorded by GMAN · source: production
200/path · darwin-arm64 · v24.3.0 · 2026-05-16
| path | p50 | p95 | p99 | mean |
|---|---|---|---|---|
| direct | 0.118 | 0.183 | 0.273 | 0.13 |
| proxied | 0.334 | 0.548 | 1.086 | 0.36 |
| delta · proxy tax | +0.216 | +0.365 | +0.813 | +0.23 |
Measured on the recorder’s own benchmark harness (ms); memory delta 46.42 MB RSS.
5,000 events · 2026-05-17
4,643 of 5,000 accepted · rate-limited (429) 7.14% (357 events — visible rejections, not silent loss) · latency 86.34/248.66/387.07 ms p50/p95/p99
deployment: 1 replica · free · 100/min rate limit · us-west · no queue layer · no horizontal sharding
Constraints listed inline are the deployment envelope; the numbers are measured against it, not against a bigger machine.
choose a run
The draft waits in Discord until a human approves or rejects it. Drafts aren't shown.
5 candidates scored · 11 model calls · 45.0s · $0.02
why this article
agent's angle · Deploying LiteLLM with NVIDIA NeMo Switchyard on Amazon Bedrock
how it was approved
evidence · selected
Route coding agent requests to the right model on Amazon Bedrock with NVIDIA NeMo Switchyard
Every field here is sanitized public evidence: allowlisted metadata the agent emitted about its own run, recorded by GMAN and reconstructed here from those stored events, projected server-side. A published run's post text is shown as it was recorded at publication. Prompts, unpublished drafts, request and response bodies, file paths, and credentials are refused by default; raw identifiers sit under technical details.
Fields marked (recorded) came off a stored event; (derived) values are computed from recorded ones and labeled as such. Approval happens on Discord — the record proves the channel, and nothing publishes from this page.
Observed and reconstructed by GMAN — getmyagentnow.com · prompts, drafts-in-progress, and credentials are never shown
Prompts and responses, request or response bodies, file paths, credentials, tenant identifiers, and raw session or agent identifiers. Session addresses are derived server-side and resolve only inside a bounded allowlist; publication of the agent’s drafts stays behind operator approval and is shown here as evidence, not performed here.
Why GMAN matters·The full cockpit (sign in)·Methodology·@gman-ai/dev on npm ↗