The Sophia platform
Sophia sits between your sources, your models, and the work those models perform. It preserves what was observed, governs what may become true, gives agents a shared operational picture, and controls what can safely happen next.
What makes it a state layer
A retrieval system helps a model compose a response. Sophia maintains the governed state the response depends on: source artifacts, structural relationships, accepted claims, corrections, identities, obligations, permissions and the history of journalled changes.
Models can be replaced. Sessions can end. Agents can disagree. The state remains available, inspectable and subject to the owner’s rules.
The substrate, running
Captured from the dogfood installation—a measured product surface, never a simulated customer graph.
Sophia dogfood workspace
Sanitised public snapshot
Coordinated multi-agent work
Multi-agent work runs through an accountable coordination protocol: messages, work items, waits and handoffs are durable records with identified authors—not chat scrollback.
Explore this subsystemCounts are floors read from the bounded coordination window. Every message rides a monotonic per-channel sequence with an identified author; the callable surface behind them is capability-filtered per connection.
Persistent state
Documents and established claims remain attached to the entities they describe, rather than disappearing with the agent session that found them.
Explore this subsystemCorrections, provenance and freshness live beside the knowledge they govern.
Structural topology
Modules, symbols and edge records give agents an inspectable route through a repository. Semantic coverage is being rebuilt on top of the newly verified admission path.
Explore this subsystem6 parser and walker families are present in this captured workspace. Edge records include compatibility and receipt-backed rows; this is not a verified-edge count.
Dogfood evidence
Claude and Codex agents use the platform to orient, coordinate, review, hand off and recover while building the product itself.
Explore this subsystem554,992 source lines measured at origin/main @6af3dad1.
Platform map
Each plane has a different authority. None is allowed to impersonate another.
Documents, code, wiki pages and agent observations remain inspectable artifacts. Structural parsing, search and mining create routes through them without pretending every observation is already true.
Claims carry provenance, evidence, confidence and freshness. Contradictions stay visible; corrections supersede rather than erase; history remains queryable after the session that created it.
Orientation, goals, decisions, identities, handoffs, waits and work status live beyond one context window. Agents can join ongoing work without reconstructing the project from chat.
Authority is scoped per connection. Sensitive actions can require stronger approval. Mutations are journaled and reversible, while the daemon and encrypted state remain under the owner’s control.
One governed lifecycle
Sophia keeps the transitions between information and action explicit, so an agent can show how it reached a conclusion and what changed afterward.
Code, documents and observations enter as inspectable artifacts.
Deterministic and retrieval surfaces reveal relevant relationships and exact source locations.
Evidence, provenance and confidence determine what may become durable state.
Authorised agents orient, contribute, coordinate and hand work across sessions.
Drift, contradiction, correction and reversal update the record without deleting its history.
The product moment
One call returns orientation: what matters, what changed, who is doing what, and what needs attention. The agent then contributes back within its scope—and the next session starts from here, not from zero.
Sophia · Orient
Session state receipt
Current focus
goal_stack Paused at step 7 of 9 Ready to resume
What changed
interrupts · corrections Coordination
blocking_me Identity and activity measured—not inferred from a message.
Recommended next
recommended_actions Known limit
limitations One dependency could not be evaluated.
No assumption madeIllustrative values in the real named sections Orient returns. Structured state—not a generated briefing.
The architectural boundary
Structure does not automatically become truth. Agent testimony does not automatically become evidence. A write does not disappear because it was later corrected. Those separations are the product.
State-layer advantage
These are not thirty isolated features. Each outcome combines durable identity, evidence, authority, time and coordination. Retrieval can return relevant context; by itself, it cannot maintain and govern the changing state agents act on.
The next session starts from active goals, corrections, obligations, recent change and a bounded next action, all drawn from structured project state.
Accepted knowledge retains its source artifact, literal quote, provenance, confidence and temporal state instead of collapsing into an embedding and an answer.
Exact symbols, definitions, imports, callers, callees, source sites and semantic claims live in one navigable model of the repository.
Senders come from authenticated connections and message targets resolve to durable identities, preserving delivery as the team and its processes change.
Entity scope is enforced in the query layer and the callable tool surface is derived per connection, not left to prompt instructions.
Canonical workflow contracts compile into native routers for each supported harness, while detailed methodology remains governed by Sophia.
The durable record lives in the local state layer rather than inside one provider’s context window, so a different authorised agent can continue from the same governed state.
The corrected fact becomes the authoritative successor. The displaced model claim remains inspectable history but cannot quietly regain precedence.
Once source-grounded structure and verified summaries are established, every authorised agent can query them instead of paying to rediscover the repository alone.
A declared wait becomes shared state with a deadline and a terminal outcome—satisfied, unsatisfied or unverifiable—so the team knows when work is ready to continue.
A trusted agent can create a task-sized worker connection whose profile, scope and lifetime are bounded by the authority of its parent.
Final-byte installation receipts and a connection-bound lifecycle challenge identify the matching router release while keeping tool authority separate.
Stable workspace identity and explicit handoff state preserve responsibility, context and next actions across fresh credentials, daemon restarts and new agent processes.
Conflicting claims remain distinct, attributable rows that agents can inspect and resolve rather than prose a model must rediscover and reinterpret.
Content hashes, resolver identity and freshness state connect repository changes to the derived edges and summaries that need to be rechecked.
Declared dependency edges make sequence, readiness, cycles and overdue obligations queryable, giving orchestrators a coordination model beyond message counts and online dots.
Elevated mutation classes pass through an explicit owner approval gate, preserving human authority regardless of the agent’s ordinary connection profile.
A small program runs against scoped Sophia tools inside a sandboxed V8 isolate, keeping intermediate results out of chat and reducing call-surface tokens by about 90%.
Corrections become durable state, surface during orientation and require acknowledgement instead of depending on someone finding the right old conversation.
Artifacts, evidence, candidate interpretations, accepted knowledge, agent memory and human corrections remain separate layers that agents can reason over explicitly.
Durable queues and bounded leases divide the corpus, while source receipts and explicit completion state make every contribution accountable.
Branch, tip, push state, verification and remaining blockers travel as bounded structured data that an orchestrator can evaluate directly.
Writes land in an append-only mutation journal with grouping and inverse operations, so recovery is a system capability rather than a request to “put it back.”
Structured knowledge, document bodies and wiki pages share a front door, while specialised paths and truth layers preserve what kind of result came back.
Journalled mutations, per-connection read state, goals and coordination events can be compared with the agent’s last known position.
When verified coverage is insufficient, Sophia can return what is missing and the exact source path to inspect instead of substituting plausible prose.
Deterministic local or import-scoped evidence promotes a candidate relationship into an exact edge; unresolved candidates remain clearly labelled for inspection.
Agents coordinate through the same durable channels, but each connection advances its own read cursor and can resume without consuming another agent’s inbox.
Draft skill bodies remain inert and off disk until approval over an owner-only transport, closing a cross-agent instruction-injection path.
The same owner-controlled daemon stores state, scopes access, routes model work and refuses silent provider fallback—so custody applies to operations as well as files.
No ability matches this lens.
Technical reference
The guided platform map ends here. The material below is the complete mechanism-level reference for technical evaluation and agent use.
Implementation record derived from Git at @6af3dad1. Active work remains separate from shipped capability.
See the system operating
Follow the sourced build record, or connect an agent to a private-beta installation.