mitigation — rotate exposed keyraised by the agent · evidence attached
agent → ratify ✕ no key to stamp with
PROPOSEDIN_PROGRESS · m.silva ✓
finds are automatic — decisions are human
feedbackskillspromptscatalogsroadmapyour model · your hardware
Reason01 / 08
One binary.
No dependency tree, no orchestration cluster, no frontend build. You hand IT one Go binary and it runs — agent loop, UI, and storage compiled into a single artifact. Boring technology, deliberately chosen.
Reason02 / 08
The server is truth.
There is no JSON API and no client state — the layer everyone expects simply isn't there. The agent streams HTML fragments over SSE; the browser shows them. The LLM's output is the interface.
Reason03 / 08
Four layers, one shape.
Solutions package. App glues. 33 domain slices in six families do the work, on infra bedrock. Every slice owns its full stack — a new capability is a new directory. Nothing else changes.
Reason04 / 08
The agent lives inside the boundary.
A workspace defines who may act and what exists. The agent gets the same ceiling a user gets — tools outside the boundary aren't refused, they're unreachable. And every step is appended to the record.
Monitor05 / 08
Sovereign by construction.
Solid runs on a server you own and run. Three ways in — data you ship, stores we query, APIs behind full guardrails — and zero ways out. When the agent touches a store it acts as you, on your own expiring token: it can never see more than you can.
Monitor06 / 08
From monolith to three planes.
Everyone starts with one binary on one server. Under load it differentiates — like an amoeba — into compute, one quack per workspace, and a lake of signed data, each scaling sideways on its own. Bare metal or VMs, no Kubernetes required — still inside your boundary, isolation enforced by the wire.
Trust07 / 08
Open agent, closed notary.
Above the line: open — fork it, extend it, resell it. Below: every bit hash-chained and anchored; every config change a git commit. And one thing the agent structurally cannot do: ratify its own decisions. Finds are automatic — decisions are human. Open data, closed proof. Trust compounds.
Trust08 / 08
The flywheel.
Every answer breeds feedback; feedback flows back into skills, prompts, catalogs — and, on the long orbit, into your model, on your hardware. Most agent frameworks are toolkits for developers. Solid is a platform for operators.
See it running.
The architecture is the pitch — but it's better live. One binary, your data, real questions.