Enterprise Agent Platform · Harness governance engine
Let AI do as much work as possible — never with unlimited power
The Enterprise Agent Platform gives companies in banking, FMCG, pharmaceuticals and beyond a runtime and governance foundation for agents. We define agents strictly: a large model is a brain; an agent is brain + hands + workflow. A model can only answer questions — an agent completes tasks, and it must run inside a controlled governance system.
Three layers, twelve components
The cognition layer thinks, the knowledge and action layer acts, and the governance layer controls. The model only proposes, tools execute under control, and Harness holds the final say.
- 1LLM coreUnderstands, reasons and generates — it may only propose
- 2Task goals & contractWhat to do, to what standard, and what is off-limits
- 3PlannerBreaks tasks into verifiable, replayable steps
- 4Memory & stateWorking memory, facts and open items managed in layers
- 5RAG retrievalA factual base of policy, regulation and cases — answers cite the source
- 6Skill packsStandard operating procedures turned into reviewable steps
- 7Tool calls via MCP / APIA controlled interface for queries, calculations and writes
- 8Workflow engineChains multi-step tasks, state transitions and human handoffs
- 9Harness runtime controlThe master gate: the model proposes, the system decides
- 10Rules enginePolicy as code — is the business action allowed?
- 11Permissions & authorizationIs there authority to act? Re-checked on every call
- 12Audit trailWhat actually happened — tamper-proof and replayable
One governance foundation, agents for every industry
Harness entry, runtime and exit controls are industry-neutral; what changes by industry are the Skill packs, tool interfaces and gate rules.
Corporate credit and other scenario agents
A nine-stage run where AI does 80% of the preparation and people keep 100% of the decision; policy Q&A and post-loan alerts.
See the solution
Campaign reviews and first-pass rebate checks
Drafts campaign reviews from the data and checks rebate and claim documents against the rules, passing exceptions to people.
See the solution
Expense checks and case preparation
First-pass compliance checks on marketing expenses and preparation of pharmacovigilance case information — every conclusion confirmed by a named person.
See the solutionFour things agents can do. Three they cannot.
Hand these to agents with confidenceFour can-dos
- Information work — retrieve, understand and explain; every answer cites a source, and no source means no answer
- Document processing — multimodal understanding that goes from “seeing text” to “understanding the file”
- Analysis & investigation — standardized legwork automated end to end, each result sourced, time-stamped and filed in a unified case
- AI Reviewer (a second pair of eyes) — never replaces the approver, but systematically finds human oversights before sign-off
Boundaries that must holdThree cannots
- No technical 100% guarantee — freedom from hallucination, always-correct reasoning, true understanding, a complete view of the world, foresight of extreme events and consistent results cannot be promised, so model output is never treated as company ground truth
- No unlimited business autonomy — every access must fall inside the intersection of eight elements: identity, role, business purpose, customer consent, task, data scope, least privilege and validity period
- High-risk decisions are never “decided by one model sentence” — credit approval, funds transfers, asset valuation, large expense settlements and drug-safety judgements must keep authorized sign-off, a human intervention channel and an emergency kill switch
The Five Inequalities
Capability, fact, decision and execution are gated and logged, layer by layer.
Six autonomy levels, L0–L5
The platform grades AI autonomy into six levels with tiered admission. For most companies, the main arena is L2 + L3 + conditional L4.
- L0Answers questions onlyPolicy Q&A, regulation lookup, product-information Q&ADeploy at scale
- L1Performs analysisFinancial analysis, cash-flow profiling, sales analysisDeploy at scale
- L2Gives recommendationsMain arenaCredit recommendations, promotion proposals, risk alertsDeploy at scale
- L3Calls toolsMain arenaCredit-bureau and registry checks, inventory lookups, drug sales-flow lookupsControlled deployment
- L4Executes low-risk tasksMain arenaDocument requests, reminders, first-pass claim alertsConditional automation
- L5Independent high-risk decisions and executionAuto-approval + auto-disbursement; auto-settling large rebatesNever “unguarded” in principle
Entry, runtime, exit: fifteen control modules
Harness is the platform’s core governance component — and the precondition for any agent to enter a production system.
- 1Task admission controllerDecides whether a case may be handled by an agent and at what autonomy; otherwise stops before the model starts
- 2Identity & authorization managerRe-checks operator, customer, data and purpose on every tool call — not once at case opening
- 3Task plannerMakes the model produce a structured plan and validates each item; invalid plans are re-planned
- 4Business state machineHolds legal case states and transitions — the model may only request changes; Harness rules
- 5Context assemblerLoads only what the current step needs and re-injects authorizations, prohibitions and amount limits every turn
- 6Case state & memory managerSplits information into working memory, facts, open items and immutable audit events
- 7Skill executorRuns procedures in order; no skipping after a failure — the sequence itself is risk control
- 8Model routerSelects and pins model versions per task; no mid-run upgrades, so results are reproducible
- 9Tool gatewayEnforces eight duties on every call, including schema checks, authorization, allow-lists, rate limits, masking and idempotency
- 10Policy gateTurns what cannot rely on model good behavior into nine deterministic policies that block before output
- 11Independent validatorRe-verifies input, process, result and action; reads only structured data, never trusts natural language
- 12Human interaction & approval managerPresents a full approval pack with evidence for and against; approvals bind to exact parameters and this case
- 13Completion sensorJudges completion by a conjunction of conditions — not by “the report has been generated”
- 14Exception recovery & compensationConfirms whether a previous write succeeded before retrying — no blind replays or duplicate writes
- 15Audit monitoring & emergency stopTraces everything from authorization to reconciliation; supports read-only degradation, human takeover and tiered shutdown
Hard constraints where model good behavior can’t be trusted
Policy gate
Requirements that cannot depend on model good behavior are written as nine deterministic policies that block before model output.
Tool gateway
Eight duties on every call: parameter schema validation, identity and authorization checks, tool allow-listing, frequency and volume limits, sensitive-field masking, return-value integrity checks, logging, and idempotency control for writes.
Task contract
What counts as done, and when to stop, become hard constraints: completion requires all ten conditions, stops are executed by triggers. No more “I think I’m finished”.
Professional Skills
Standard operating procedures become steps that cannot be skipped. Every metric is computed deterministically; the model only explains and suggests what to check.
- The modelproposes “what I want to do”
- Harnessdecides “can it be done?”
- Rulesdecide “is it allowed?”
- Permissionsdecide “is there authority?”
- Auditrecords “what actually happened”
The agent paradigm, first applied to modeling
We were first to bring the agent paradigm to SAM: one supervisor, seven specialists, three cross-step roles, fourteen named human gates — and a Release Controller as the single gate to production.
Watch a high-risk scenario run (banking)
The corporate credit agent’s nine-stage run shows “AI does 80% of the preparation, people keep 100% of the decision” in full.
Start with one business question. Let SAM deliver a model you can sign off.
Whether you are in banking, FMCG and marketing, pharmaceuticals or another industry, try SAM online first — or talk to our modeling experts about your use case.