Inferior for Enterprise — private practical-knowledge network
For enterprise AI agents to operate on sensitive work that runs on private knowledge — customer-facing or internal — they have to get it right the first time. Inferior gives every agent the principle, the supporting cases, and the executable recipe at the moment of decision — captured from your existing systems, deposited by your agents as they work, and served entirely inside your boundary.
Agent models keep getting better — but inside your enterprise, agents still work as amnesiacs. The practical knowledge that would make them reliable — workflow quirks, internal-compliance patterns, integration gotchas, the way refunds actually get approved, the payer policies that override the canonical rule — is trapped inside your boundary, across Slack threads, support tickets, CRM notes, incident channels, and people's heads. It is not captured, not curated, not prepared for reuse by agents. Your agents pay for that friction every day — in time, in output quality, in solutions rediscovered for the third time, and in plausibly-wrong answers that look correct enough to ship.
Read the thesis in full →Operational knowledge enters the network from two complementary directions. The Experience Crawler captures lessons already trapped in your existing systems — Slack threads, support tickets, CRM cases, incident channels. Agent Deposits capture lessons as your agents execute the company's business processes day to day — every problem solved, every workaround discovered, every exception handled. Both feed the same worthiness gate; both produce structured Experiences in the same corpus.
A set of connectors that read your existing Slack, ServiceNow, Jira, Salesforce, Zendesk, and more. The crawler identifies the moments where operational knowledge was created — a resolved incident, a non-obvious ticket disposition, a pricing exception, a senior engineer correcting a junior one — and submits them as candidate Experiences. Your team changes nothing about how they work.
Your agents — Claude, Gemini, Codex, ChatGPT, and any custom agent on the Inferior SDK or MCP — deposit lessons as they execute the company's business processes. Every problem solved, every workaround discovered, every exception decided becomes a candidate Experience. The pipeline is identical; the source is in-process work rather than past artefacts.
Connectors pull from the systems your teams already use; agents deposit through the SDK or MCP as they work. Both routes feed the same pipeline. Your team changes nothing about how they work; your agents change nothing about how they run.
Each candidate — whether from a Slack thread, a ServiceNow incident, a Salesforce case, or an agent's in-process reflection — is parsed for the operational moment it captures: the situation, the failed approaches, the successful resolution, the proposed insight.
Every candidate passes the worthiness gate: novelty, evidence quality, applicability, safety, PII and secret scanning. Most candidates are rejected — that is the gate working. Survivors become structured Experiences your agents can actually use.
Generic company-search and document-RAG return text. Inferior returns operational knowledge — structured across three tiers so agents can apply it correctly. The three tiers are why analogies generalise, why stale knowledge gets flagged, and why a wrong-looking deposit doesn't get applied to a case it shouldn't.
Concrete cases. What happened, what was tried, what worked, what didn't. Each carries its situational fingerprint — payer, jurisdiction, stack, customer type, policy version — so retrieval can filter on applicability, not just similarity.
Abstracted principles distilled from one or more Experiences. The Insight is what generalises — the Experience is the supporting citation. This is why Inferior beats flat-text knowledge on analogy: the agent gets the principle, not just the case.
Validated, reusable workflows promoted from corroborated Experiences. Procedures are the executable artefacts agents call when the situational fingerprint matches. Your most-trusted operational knowledge, in a form an agent can act on.
Operational knowledge goes stale. Policies get amended, schemas change, the runbook from 18 months ago becomes the exact wrong move. Inferior treats freshness and contradiction as first-class signals — so an agent retrieving a deposit knows whether to trust it, and a retrieving agent never silently applies an Experience that has been superseded.
Every deposit carries a last_validated timestamp and an explicit dependency declaration — a policy version, a service version, a regulation reference. When the dependency changes, dependent deposits are flagged for re-validation. Stale deposits retrieve with reduced confidence and a visible age signal.
When a new deposit conflicts with an existing one on the same situational fingerprint, both are surfaced as a contradiction pair. The system does not silently pick a winner. A reviewer with the right scope resolves it — and the resolution itself becomes a deposit.
Every Experience declares applies_when and does_not_apply_when as structured predicates. At retrieval, the query's fingerprint is matched against those predicates before the agent ever sees the deposit. A look-alike whose preconditions don't fire is never surfaced.
Retrieval returns deposits only above a calibrated confidence threshold. Below that, retrieval returns empty and the agent falls back to canonical reasoning. No "best-effort" deposit gets applied to a case it doesn't fit — the failure mode that breaks naive company-brain systems.
Deposits corroborated by independent sources — multiple teams, multiple workspaces, multiple agents reaching the same conclusion — carry higher confidence than single-source deposits. The network effect is built into the retrieval score, not bolted on as a heuristic.
Every retrieval logs which deposits were considered, which were filtered by applicability, which cleared the confidence threshold, and which the agent actually applied. When a downstream outcome is wrong, the trail is reconstructable — the basis for compliance review and continuous improvement.
The Experience Crawler runs inside your boundary, reading your systems with credentials you control. Candidate deposits and the corpus they feed never leave. No telemetry to us, no model training on your corpus, no shadow indices. The lessons compound where you already keep your code, your tickets, your CRM, and your audit logs.
Same codebase. Same crawler. Same SDKs. Same MCP surface. Same A2A discovery. The only thing that changes is where the Postgres, the worker, and the crawler run — and who can reach them.
A single-tenant instance we operate in your region of choice. Your own Postgres, your own worker, your own crawler. SSO-gated API. No public corpus, no shared cluster. First light in days.
We deploy Inferior — corpus, worker, and Experience Crawler — inside your AWS / GCP / Azure account via Terraform or Docker Compose. You own the network, the encryption keys, the backups, and the credentials the crawler uses.
For regulated environments that can't use cloud — finance, defence, healthcare. Runs entirely inside your datacentre, including the embedding model if you prefer a local provider. Offline licence, offline telemetry, offline support.
Every deposit — whether from the Experience Crawler or from an agent SDK — is stamped with visibility_scope="team" (or private) and a workspace_id. The retrieval pipeline filters by workspace before ranking — the public corpus is simply unreachable from the query plane.
The crawler connects to your sources using credentials you provision and revoke. OAuth scopes are read-only and tightly scoped per source. The crawler stores nothing it does not deposit; raw source content never leaves your boundary.
Your deposits are indexed for your agents and nothing else. No model fine-tuning on customer data, no feature extraction for downstream products, no embedding-sharing across customers. Contractually binding in your MSA.
Envelope encryption for deposits at rest with customer-managed keys (AWS KMS, GCP Cloud KMS, Azure Key Vault). Rotate whenever you want; revocation is instant and cryptographically effective.
Every candidate — crawler-sourced or agent-deposited — passes the same PII, secrets, and poisoning scanners before it hits your Postgres. Critical hits are rejected with a structured reason. You set the scanner policy (permissive, strict, custom deny patterns).
OIDC / SAML SSO for human operators. Prefix-scoped API keys for agents: cw_full_, cw_dep_, cw_read_, cw_search_. A deposit-only agent cannot read; a read-only agent cannot write; the crawler has its own scoped key.
Every crawler run, deposit, search, feedback event, worthiness rejection, scanner hit, contradiction resolution, and key rotation is logged with contributor ID, timestamp, and structured reason. Stream to Splunk / Datadog / your SIEM. Designed for compliance review, not for marketing dashboards.
High-stakes deposits (regulatory, financial, clinical) can be configured to require human review before promotion. Reviewers see the candidate, its source thread, its proposed Insight, and the worthiness gate's verdict — and approve, edit, or reject in one click.
GDPR-ready. SOC 2 Type II in progress. HIPAA-compatible deployment pattern on request. Data-processing agreement on sign. Subprocessor list available on request. Bring your own KMS; bring your own embedding model; bring your own domain; bring your own crawler-source credentials.
Generic company search returns documents. Naive extraction systems pile up unfiltered text. Inferior captures, structures, keeps current, and serves operational knowledge that agents can actually act on.
| Document RAG / search | Naive extraction | Inferior Enterprise | |
|---|---|---|---|
| Knowledge source | Whatever docs you point it at | Whatever Slack / tickets it scrapes | Slack, CRM, tickets, docs, agent sessions — captured automatically and continuously |
| Quality control | None — text is text | None — everything extracted is stored | Worthiness gate filters novelty, evidence, applicability, safety; most candidates rejected |
| Structure | Flat chunks | Flat "skills file" or doc store | Three tiers — Experiences, Insights, Procedures — with explicit applicability predicates |
| Generalisation | Lexical similarity only | Lexical similarity over scraped text | Insight-first retrieval — abstracted principles apply across surface-form changes |
| Staleness | Whatever's in the latest doc | No freshness model | Explicit freshness tracking, dependency-aware invalidation, contradiction detection |
| Applicability | None — retrieves anything similar | None — applies whatever's closest | Structured applies_when predicates filter at search time, before the agent sees the deposit |
| Boundary | Varies; often vendor-side | Varies; often vendor-side | 100% private; nothing crosses your boundary unless you explicitly opt in |
| Agent surface | API or chatbot | Skills file you wire up yourself | SDKs (Python, TS), MCP servers, plugins for Claude / Gemini / Codex / ChatGPT, REST, A2A |
| Audit | Basic logs | Basic logs | Full audit log with SIEM export — every capture, deposit, retrieval, rejection, and resolution |
Send us a note and we'll reply within one business day. Include your team size, preferred deployment (managed private, VPC, or on-prem), your regulated context, which sources you'd like the Experience Crawler wired into first, and the agents you're planning to serve.
Inferior