Antimetal · Production research

Production that runs itself.

It reads the live graph without waiting for a ticket.

It acts on the runtime with a sealed, reversible plan.

It improves the next loop from what last broke.

FIG. 01 · WORLD GRAPH

The agents already knew the blast radius before we opened the ticket. That is the difference between a dashboard and a colleague.

Maya Chen

Staff SRE

Google

FIG. 04 · INCIDENT HALF-LIFE

30d · n=418

AntimetalBaseline ops

Aproduction graph is not a status page. It is a moving mechanical drawing: lag, certificates, queues, and the people who still have to sign.

Human-in-loop work decays slowly. Autonomous close, when the world model is current, drops the half-life of an incident instead of decorating it.

We keep the drawing thin on purpose. Color is reserved for the one curve that actually bent.

A new layer of the stack

Manifesto

Production should run itself.

FIG. 07 · CONTROL PLANE

01 Your Team

intent · policy · review

02 Antimetal Agents

patrol · triage · repair

03 Antimetal World Model

graph · memory · blast radius

ACTIVE

04 Cloud platforms

AWS · GCP · Azure · k8s

05 Production

runtime that stays alive

AWS
EKS
RDS
IAM

02 · Runtime

A layer that owns the runtime.

Observability watches. Orchestration schedules. Antimetal sits between your team and the cloud and keeps a current model of what is allowed to change. Agents do not chat about production; they hold it.

04 · Operation

Everyone else watches. We operate.

A ticket is a delayed drawing of something the graph already knew. The useful layer is the one that can freeze writes, rotate a cert, and leave a record a person can still read at 03:00.

FIG. 12 · WORKING DESK

Console · prod-us-west

Runtime desk

LIVE

Health

93%

Response

2m 12s

Uptime

99.97%

Pending approval

Scale redis-prod west-2

LOW

Replica lag at 840ms. Propose +2 nodes, freeze writes 12s.

ApproveHold

Rotate edge cert edge-17

MED

Expiry in 19h. Canary on 4% traffic, rollback if TLS error > 0.2%.

ApproveHold

Event log

  • 02:14:08world-model reconciled 1,204 nodes
  • 02:14:21patrol flagged p95 drift on checkout-api
  • 02:14:44triage assigned blast radius: payments, auth
  • 02:15:03agent-builder drafted runbook AM-4412
  • 02:15:18awaiting human seal on redis-prod scale

One platform. An army of agents.

Four loops share one world model. Switch the row to see the trace they leave, not a blank frame.

PATROL LOOP

LIVE TRACE

Continuous surface scan

  • 02:14:08 scan checkout-api p95 +18%
  • 02:14:11 isolate redis-prod west-2
  • 02:14:13 anomaly locked node 274

Questions, filed thinly.

No. It takes the first hours of recognition and repair. People still set policy, approve irreversible work, and own the product.

From the research log.

LOG 19

A mechanical model of replica lag

Treating cache topology as a linkage, not a dashboard, so agents can feel load transfer before alerts fire.

I

L. Ibarra

Research

LOG 24

Terrain maps for multi-region failovers

A three-dimensional mesh of latency ridges. The path of least regret is rarely the shortest line.

O

S. Okonkwo

Systems

LOG 31

Unrolling the stream of a live incident

Field notes on drawing the event as expanding streamlines instead of a linear timeline.

V

A. Voss

Design

The autonomous system for production.

A thin control layer for teams who would rather keep the runtime than decorate another status board.

Product

Company

Notes

Contact

System nominal
From We0