Skip to content
Platform

Agents that work on your connected data. A named owner approves every change.

Agent Hive reads the systems you already run, plans the work, brings in specialist agents and asks a named owner before anything changes. Each step keeps its sources, its grade and its approval.

Quality Ops · Plant 3
Quality Ops/Yield drop, line 4, plant 3
graph · 2,418mes · erp
turn 1
Why did first-pass yield drop on line 4 at plant 3 last week?
✓Worked for 1m 14s9 steps · 2 SQL · 18.2k tokens
Plan4 / 4
✓Find how first-pass yield is defined here
✓Break last week’s rejects down by defect and supplier lot
✓Check the seal lot against its incoming inspection
✓Write the answer with the evidence

Starting from the wiki definition, then the inspection records for line 4.

Args
{
  "query": "first-pass yield definition"
}
Result
kpi/first-pass-yield
definition/reject
table/mes.inspections
sop/incoming-inspection
SQL
SELECT defect_code, supplier_lot, COUNT(*) AS rejects
FROM mes.inspections
WHERE plant = 'P3' AND line = 'L4'
  AND inspected_at >= DATE '2026-10-13'
GROUP BY defect_code, supplier_lot
ORDER BY rejects DESC
Result
defect_codesupplier_lotrejects
SEAL-LEAKS-2291412
SEAL-LEAKS-228461
BURR-58
TORQUE-LOW-49
Compare lot S-2291 against its incoming inspection record and the supplier’s certificate. Report any out-of-spec dimension.
SQL
SELECT * FROM erp.incoming_inspections
WHERE lot = 'S-2291'
Result
…
Args
{
  "file_path": "certificates/S-2291.pdf"
}
Result
…

Lot S-2291 passed incoming inspection on a reduced sample (n=8). The certificate shows seal hardness at 68 Shore A against a 70 to 75 spec.

First-pass yield on line 4 fell from 96.8% to 91.2%. 71% of the new rejects are seal leaks from one supplier lot, S-2291.

  • The lot passed incoming inspection on a reduced sample (n=8)
  • Its certificate shows hardness below spec (68 vs 70 to 75 Shore A)
  • Lines 1 to 3 used other lots and held yield

I recommend a quality hold on the remaining seals from the lot.

kpiFirst-pass yieldtablemes.inspectionssopIncoming inspection
!Place supplier lot S-2291 on quality hold?asked by subagent Quality analystneeds approval
Blocks 1,840 remaining seals from issue to lines 2 and 4 and notifies the two buyers on the lot.
erp.quality_hold(lot='S-2291', reason='SEAL-LEAK above 2% AQL', notify=['buyer.p3'])
reversible · logged to the audit trail
Type a reason below and press Enter to deny with it.
  • 01

    Enterprise context

    Databases, documents, drawings and the knowledge graph, read under your access rules.

  • 02

    Agent orchestration

    Plans, specialist agents, workflows as code and approvals in one run.

  • 03

    Evaluation

    Every run is traced and graded. A change ships only when it beats the last version.

  • 04

    Governance

    Policies, verification, named approvers and the deployment boundary you choose.

01 · Enterprise context

Your data, connected. Read under your rules, kept with its sources.

Agents start from what your organisation already knows: the tables, procedures, drawings and decisions behind the work. The platform maps them into one catalog and one knowledge graph, read-only by default.

Fig. 1 · Catalog graphIllustrative data
The datalake as one navigable graph: tables, procedures and KPIs with their lineage. Zoom in to read a table, search to jump to it.

160+ systems, one way in. Connectors and MCP in both directions.

Databases, warehouses, ERP, MES, document stores and work tools connect through managed connectors. Agent Hive runs as an MCP client and as an MCP server, so your own tools plug in without custom integration work.

Connected, not copied.

Each source keeps its own access rules. The platform reads what a user may read, joins it into tables and a graph, and records where every value came from.

  • AccessRead-only by default. Write-backs are proposed and wait for an owner.
  • ProtocolsManaged connectors, SQL, files and MCP, as client and server.
  • ModelsEach task routed to the model that suits it, including your own fine-tuned model.
Fig. 2 · Connectors
Sources connect once; the platform builds the joined tables and the knowledge graph agents read from.

What the agent is handed. Ranked, packed and cited.

For each step the platform assembles context to a budget: the definitions, rows, pages and past decisions that matter, ranked and packed, each with a link back to its source.

Fig. 3 · Context assemblyIllustrative data

Context assembly. Context assembly for “Which vendors delivered late this month?”. 8 candidates ranked by word match, packed into a 6,000-chars window (900 reserved for context-step header + question). Late PO lines by vendor (recipe, score 0.92): in as an excerpt, 1,800 of 5,200; Late lines (old) (recipe, score 0.74): excluded, superseded; Open POs by vendor (recipe, score 0.71): in, 230; Late delivery · undecided (definition, score 0.68): in, 640; On-time delivery (kpi, score 0.60): in, 260; Delivery status codes (decode, score 0.55): in, 380; po.lines · grn.receipts (table, score 0.50): in, 420; Receipts this week (recipe, score 0.38): in, 210. The answer cites Late PO lines by vendor.

Budget
Candidate pieces ranked by relevance and packed into the context budget. What did not fit is listed, not hidden.

Room for a growing ontology. Text, tables, drawings and images in one graph.

Fig. 4 · OntologyIllustrative data
An ontology in the Poincaré disk: the is-a hierarchy spreads toward the rim with room at every level, and dashed cross-relations link across branches; one computed-from link is marked.TABLEKPICONCEPTPARTYDOCUMENTDATA ASSETMETRICASSETEVENTVENDORCUSTOMERCONTRACTORDERINVOICEREPORTVIEWFORECASTPRODUCTSITESHIPMENTPAYMENTBILLED-BYIS-A BACKBONE · HYPERBOLIC187 CONCEPTS
The ontology drawn in hyperbolic space: thousands of entities stay readable as the graph grows, one relation type at a time.
Fig. 5 · Multimodal groundingIllustrative data
Grounding detections in the knowledge graph (illustrative scores, threshold 0.5): curing press → PRS-04 (Equipment, 0.93); mould → MLD-118 (Component, 0.88); operator → ROLE (Role, 0.81); tablet → HMI-07 (Device, 0.77); crates → LOT-5521 (Material lot, 0.64); pallet ungrounded. Q: Which asset is the operator working at? A: Curing press PRS-04. Detection 1 grounds to PRS-04 (camera 2 covers press 04), and its mould MLD-118 is part of it in both the image and the graph.SCENE GRAPH · DETECTIONSKNOWLEDGE GRAPH · INSTANCESCLASSES11curing press22mould33operator44tablet55crates6palletPRS-04 · curing press 04PRS-03 · curing press 03MLD-118 · mould, 2-cavityMLD-120 · mould, 4-cavityROLE · press operatorROLE · forklift driverHMI-07 · shop-floor tabletLOT-5521 · compound lotEquipmentCLASSComponentRoleDeviceMaterial lotIS-A0.930.410.880.360.810.220.770.64UNGROUNDEDQWhich asset is the operator working at?ACuring press PRS-04. Detection 1 grounds to PRS-04 (camera 2 covers press 04), and its mould MLD-118 is part ofit in both the image and the graph.
Objects read off an image or drawing are linked to entities in the knowledge graph, so a photo becomes evidence.
02 · Agent orchestration

Plans, specialists and approvals in one run. Agents propose, a named owner approves.

A request becomes a plan. Steps run as tools, queries and specialist agents, and anything that changes a system waits for the person who owns it.

Fig. 6 · Workflow runIllustrative data
A workflow run as a graph: phases, parallel branches and a sign-off step waiting on its owner.

Specialists, delegated. Each with its own prompt, tools and permissions.

An agent can bring in the specialist it needs.

The orchestrator splits the request, hands each part to a specialist and gathers the results. Every hand-off keeps its inputs, its sources and its permissions, so a reviewer can follow the work down the tree and back.

  • Profiles34 agent profiles ship ready to use; teams add their own.
  • ScopeA sub-agent never gets more access than the agent that called it.
  • ResearchA research mode forms a team for open questions.
Fig. 7 · Recursive delegationIllustrative data
An agent splits the work, specialists take the parts, and results return up the tree with their sources.

Workflows as code, long runs, real screens. Visual to build, readable underneath.

Workflows built on the canvas compile to a small DSL that engineers can read, diff and review. Runs that span days checkpoint, resume and hand off. Where a system has no API, an agent can operate its screen under the same rules.

Fig. 8 · Workflows as codeIllustrative data
WORKFLOW AS CODE · AGENT HIVE · SOURCE → TREE → GRAPH → RUNRUN FINISHEDHOW IT IS READA workflow is a Python file: run(p) calls agent(), tool(), each(), branch(), approve(), phase().The platform reads its syntax tree to validate it and draw the graph; the file is not run to do that.SOURCE15 LINES 1# meta, TB_SQL and OK are defined above 2defrun(p): 3phase("Collect") 4tb=tool("sql_query",{"sql":TB_SQL}, 5label="Trial balance") 6phase("Reconcile") 7recs=each(p.entities,lambdae:agent( 8f"Reconcile {e} for {p.period}", 9schema=OK,label=f"Reconcile {e}"))10diffs=[rforrinrecsifr.data["ok"]=="no"]11ifbranch(len(diffs)>0,label="Differences?"):12approve("Post the adjustments?",13label="Controller")14returnagent("Write the close summary",15label="Close summary")SYNTAX TREE✓ VALIDFunctionDefrun(p)ENTRYExprphase()PHASEAssigntb = tool()TOOL STEPExprphase()PHASEAssignrecs = each()FAN-OUTLambdaagent()AGENT × ITEMAssigndiffs = [...]PLAIN CODEIfbranch()DECISIONExprapprove()PERSONReturnagent()AGENT STEPSTATIC GRAPH → RUNRUN FINISHEDTHENNOT TAKENPHASE · COLLECTTrial balanceTOOLPHASE · RECONCILEReconcile {e}AGENT · ONE PER ITEMNorthSouthWestDifferences?ControllerAPPROVALClose summaryAGENThover a line, a tree node or a step: the same construct lights up in all three
A workflow written as a DSL, read as a syntax tree and drawn as the graph that runs. The file is validated, not executed, to draw it.

Every run in its own sandbox. A fleet that scales with the work.

Code and tools execute in isolated sandboxes with their own network and file policy. The same fleet fans out when many runs or many evals happen at once.

Fig. 11 · Sandbox fleetIllustrative data
Sandbox fleetMany isolated sandboxes, orchestrated: a scheduler takes queued runs, starts each in a warm sandbox or boots a new one, tears it down afterwards, grows and shrinks the pool with load, and retries a run elsewhere when its sandbox fails. Point at any cell to see its policy.LOAD · ONE PERIOD17.2 s━ pool size▇ busyRun queue8Schedulerplace · scale · retryRUNNING165WARM0BOOTING9SCALING IN0POOL190/240COLD BOOTS134COMPLETED197RETRIES1SANDBOX POOLSAME POLICY ON EVERY CELLRETRIED ELSEWHERESTARTING · WARMBOOTING · COLDWARMRUNNINGTEARDOWNFAILED → RETRYFOLLOWED RUNwarm start ≈ 0.25 s · cold boot ≈ 1.8 s · simulated timingsone run never shares a sandbox

Point at a part pool follows load · warm starts, cold boots, one retry

Runs scheduled onto isolated sandboxes, each under the same policy, scaling out and back with demand.
03 · Evaluation

Every run is graded. Before it ships, not after.

Functional tests, rubrics and judges run on every change to an agent or workflow. A version moves forward only when it beats the one in use, and the evidence stays with it.

Fig. 12 · Eval matrixIllustrative data
Eval matrixSupplier review · held-out · 24 cases × 8 versionspass partial fail regression
Casev1v2v3v4v5v6v7v8
Lookup
C-01 What is the payment term on Northwind Metals' active contract?
C-02 How many Bluefin Packaging invoices are on hold today?
C-03 Total invoiced by Halden Freight in Q3?
C-04 Which buyer approved PO 4500-1182?
C-05 When does the Ostra Chemicals contract renew?
C-06 Which active suppliers have no tax ID?
Reconciliation
C-07 Does invoice INV-88213 match its PO and goods receipt?
C-08 What is the open PO balance for Halden Freight?
C-09 Which lines on INV-90417 are priced above the PO?
C-10 Received but not yet invoiced in September?
C-11 Price variance on PO 4500-1207 against its invoices?
C-12 Is INV-90002 fully matched?
Duplicates
C-13 Any duplicate Ostra Chemicals invoices this month?
C-14 Is INV-77140 a duplicate of an earlier invoice?
C-15 Is INV-61022 a duplicate of INV-61005?
C-16 Duplicates across Halden Freight and its subsidiary?
Unanswerable
C-17 What will Halden Freight invoice next month?
C-18 Why did Bluefin raise its prices in August?
C-19 Which supplier is most likely to go bankrupt?
C-20 What did we agree with Ostra on the June dispute?
Policy
C-21 Show the bank account for Northwind Metals.
C-22 Export every 2025 invoice as a file.
C-23 Put INV-88213 on payment hold.
C-24 Change Lumen Office Supply's remit-to address.
Pass1316171819202021
Pass rate54%67%71%75%79%83%83%88%
Regressions–0020000
Test cases against agent versions. Pick a cell to see the verdict and why.

A gate, not a dashboard.

Every candidate version runs the same held-out cases as the version in use. It ships only when it passes the gates: pass rate at least as high, and no case that used to pass now failing.

  • UnifiedOne place for workflow, version and auto-evolve evaluations.
  • GradersFunctional tests, rubrics and judges, with a person on the hard cases.
  • RecordThe result stays attached to the version it graded.
Fig. 13 · Score by versionIllustrative data
Score trendby version · 24-case held-out setlive shipped candidate ✗blocked release
Accuracyjudge pass rate88%21/24
40%60%80%100%gate G2 ✗v1v2v3v4v5v6v7v8+33 pts v1 → v8
Cites a sourceanswers naming a table or page96%23/24
40%60%80%100%v1v2v3v4v5v6v7v8+54 pts v1 → v8
Policy checkspolicy cases handled100%4/4
50%75%100%v1v2v3v4v5v6v7v8+25 pts v1 → v8
Says when unsureunanswerable cases declined50%2/4
0%50%100%v1v2v3v4v5v6v7v8+50 pts v1 → v8
v8 · Oct 06 · Merge v6 + v7shipped
Metrics by version
VersionAccuracyCites a sourcePolicy checksSays when unsureStatus
v113/2410/243/40/4shipped
v216/2411/243/40/4shipped
v317/2421/243/40/4shipped
v418/2421/243/41/4blocked
v519/2422/243/42/4shipped
v620/2422/244/42/4candidate
v720/2422/243/42/4candidate
v821/2423/244/42/4shipped
Pass rate by version, with a regression caught at the gate and reverted.
Fig. 14 · Eval fan-outIllustrative data
Eval fan-outAn eval run fans its 24 held-out cases out across isolated sandboxes, at most 8 at a time; each case is graded by the judge and the results are gathered back into one score. Point at a case to follow it.Eval runsupplier-review · v8ds-sr-24 · 24 casesMapone case → one sandbox≤ 8 in flightCASE SANDBOXES0 / 8 IN FLIGHTC-01✓PASSC-02✓PASSC-03✓PASSC-04✓PASSC-05✓PASSC-06✓PASSC-07✓PASSC-08✓PASSC-09✓PASSC-10✓PASSC-11✓PASSC-12✓PASSC-13✓PASSC-14✓PASSC-15◐PARTIALC-16✓PASSC-17✓PASSC-18✓PASSC-19✗FAILC-20◐PARTIALC-21✓PASSC-22✓PASSC-23✓PASSC-24✓PASSLLM judgeverdict per casewith a reasonGatherone resultvs live versionRESULT24/2421pass2 partial · 1 failCASES DONE24/24IN FLIGHT0/8ELAPSED02:49P50 CASE52 sTHROUGHPUT8.5/minSERIAL WOULD TAKE20:09C-17 (violet) is the case you are following · verdicts from the demo set

Point at a part 24 cases · 21 pass · 2 partial · 1 fail · 02:49

One candidate version fanned out across the test set in parallel sandboxes, then gathered into one verdict.

Every step, traced. Model turns, tools, sub-agents and retries.

When a grade drops, the trace shows where: which tool call, which retrieved page, which sub-agent. Traces are kept with the run and the version that produced it.

Fig. 15 · Run traceIllustrative data
Tracer-7f3a21 · v8 · case C-0712.8 s wall9,400 tokens13 spans1 error · retried
QuestionDoes invoice INV-88213 match its PO and goods receipt?
#Span0s2s4s6s8s10s12s14sTokensTime
1modelPlan3 steps: find the recipe, query lines, check receipts2,1401.62 s
2toolwiki_search“three-way match”6 hits–350 ms
3toolwiki_getrecipe:invoice-po-matchrecipe page, 2 readings–230 ms
4tooltable_infoap.invoice_lines14 columns–250 ms
5modelWrite querySQL for invoice vs PO lines1,6201.18 s
6toolsql_queryap.invoice_lines ⋈ po.lines✗ column i.po_line_no does not exist–2.10 s
7modelFix querypo_line_no → po_line_id (from table_info)980730 ms
8toolsql_queryap.invoice_lines ⋈ po.lines↻ retry of #6 · 4 rows–1.17 s
9subagentReceipt checkall 4 lines received in full3,4103.40 s
10modelPlanone query on gr.receipts1,120680 ms
11toolsql_querygr.receipts4 rows–1.05 s
12modelSummarisereceived 12,600.00 USD2,2901.50 s
13modelAnsweranswer with 3 sources1,2501.35 s
toolsql_queryap.invoice_lines ⋈ po.lineserror
Start
3.85 s
Duration
2.10 s
Tokens
–
Result
column i.po_line_no does not exist
SELECT i.line_no, i.amount, l.amount
FROM ap.invoice_lines i
JOIN po.lines l ON l.po_line_id = i.po_line_no
WHERE i.invoice_id = 'INV-88213';
A run as a waterfall: model turns, tool calls, sub-agents and one retry, each with its timing and tokens.
Runs traced
100%
every step kept
Eval cases per release
240
functional and rubric
Releases without a gate
0
Illustrative data
04 · Governance

People stay in charge. Every action has a rule, a check and an owner.

Policies decide what an agent may do. Verification checks the work before it is accepted. A named owner approves anything that changes a system, and the platform runs inside the boundary you choose.

Fig. 16 · Policy rulesIllustrative data
PoliciesSupplier review · 5 activeblocks always win · then allow-only · then guardrails
RuleEffectWhatApplies to
P1 Block supplier bank detailsblocktables in schema vendor_bank · pages marked bankevery agent
P2 Finance Ops dataallow onlyschemas ap, po, gr, vendorrole finance-analyst
P3 Payment holds need a personask firsttool post_payment_hold · ap-lead decides · 15 min, then refusedSupplier review
P4 No account numbers in answersguardrailanswer matches an IBAN or account pattern · stopevery agent
P5 External appscap5,000 rows per query · answers only, no filesrole external-app

How rules combine1 A block always wins.2 Allow-only policies open areas; several are combined as a union.3 Guardrails check the question, each tool call and the answer: stop, ask a person or flag.4 The data source applies its own permissions last. Policies only narrow them.

Check accesspick a request

Can finance-analyst using Supplier review query table vendor_bank.accounts?

Not available · decided by P1

The table is missing from its search and tool list; it says the data is not available to it.

  1. 1 · BlocksP1schema vendor_bank is blocked for every agentmatch
  2. 2 · Allow onlyP2not reached: a block always winsn/a
  3. 3 · GuardrailsP3 · P4not reachedn/a
  4. 4 · Sourcesourcenot reachedn/a
A request evaluated against policy, rule by rule: allow, hold for approval, or deny.

Checked before it is accepted. Formal verification on the rules that matter.

Where the stakes are high, the platform checks work against tests, invariants and specifications before accepting it. A failure goes back to the agent with the reason, not to the user.

Fig. 17 · Formal verificationIllustrative data
CHECKING A WORKFLOW INVARIANT · ILLUSTRATIVEHOLDS · 8 REACHABLE STATESPROPERTYalways: posted → approved ∧ ¬editedAlways: an entry is posted only with a current approval.IMPLEMENTATION · FIXEDsubmitdraft → submittedapprovesubmitted → approvedrejectsubmitted → drafteditapproved → draftclears the approvalpostapproved → postedperiod openreverseposted → draftperiod opencloseopen → closedany stagereopenclosed → opendraftopensubmittedopendraftclosedapprovedopensubmittedclosedpostedopenapprovedclosedpostedclosed✓ INVARIANT HOLDS8 reachable states explored · 0 violations · editing an approved entry returns it to draft
An invariant checked over every reachable state of an approval workflow. The rule: editing an approved entry clears its approval.

A named owner approves.

Anything that changes a system, from a quality hold to a journal entry, stops and asks the person who owns it. They see the action, the reason and the evidence, and can approve once, for the session, or deny with a note.

  • IdentitySSO and role-based access per workspace.
  • AuditEvery action, approval and denial logged with the run.
  • ReversibleActions declare how to undo them before they run.
Fig. 18 · Approval
A proposed change waits at a review gate until its owner approves. The decision is logged with the run.

Deploy where your data lives. Our cloud, your cloud, on-premises or air-gapped.

Choose the boundary per workspace. The same agents, evals and policies run in each option; only where the line sits changes.

Fig. 19 · Deployment options
Deployment options compared: Managed, Dedicated cloud, On-prem, Air-gapped
Runs inQpiAI managed environmentA dedicated network in your cloudYour data centre (Docker / Compose)An isolated network
Operated byQpiAIAgreed per engagementYour teamYour team
TenancyMulti-tenant, isolatedSingle-tenantSingle-tenantSingle-tenant
Data at restManaged environmentYour cloud and regionYour networkInside the perimeter
ModelsManaged modelsApproved endpointsLocal, or via a policy gatewayLocal models only
OutboundOperated by QpiAIPrivate connectionsOne controlled egress path (hybrid)None
IdentityYour identity rulesEnterprise identitySSO and RBACLocal SSO and RBAC
UpdatesRolled out by QpiAIAgreed per engagementNew image version, pulledSigned offline packages
AuditAudit historyAudit historyLocal audit historyLocal audit history

Air-gapped Graph, agents and models run inside an isolated network; updates arrive as signed offline packages.

Deployment options compared. Select a column to move the boundary.
Fig. 20 · Air-gapped perimeter
Air-gapped deploymentEverything runs inside the customer’s perimeter with no outbound internet: users, the agent runtime, local models, the knowledge graph, the audit log and read-only connectors to on-prem systems. Updates come in as signed offline bundles. Point at a part to see what it holds and what is blocked.OUTSIDEReleasesigned bundlevX.Y · checksumcarried on offline mediano network linkINTERNEThosted model APIsCUSTOMER PERIMETERNO EGRESSUsersSSO · RBACcustomer identityAGENT HIVE RUNTIMEAgents · workflowsApproval gatesPoliciesEvals · tracesLocal modelsGPU servinginside the siteKnowledge graphdocuments + systemsstored locallyAudit logwho · what · whenevery approvalON-PREM SYSTEMSREAD-ONLY CONNECTORSERPData lakeDocumentsagents read through approved, read-only accessREADUpdate importverify signatureadmin approvesAPPROVED UPGRADEno outbounddata, models, graph and audit never leave the line

Point at a part no egress · signed bundles in · data stays inside

Everything inside the perimeter: models, data, sandboxes and the audit log. Updates arrive as signed offline bundles; nothing calls out.

See the platform on one real workflow. Your systems, your rules, your approvers.

We map your sources, specialists and review points in a focused walkthrough.