We use cookies to personalize content and to provide you with an improved user experience. By continuing to browse this site you consent to the use of cookies.
Please visit our Cookie Policy for further details
AI Agent Studio
Agent Studio is where business operators and engineers configure specialized agents that understand enterprise context, use approved tools, collaborate with other agents, and execute multi-step workflows with human oversight where it matters — on the dSilo agent runtime, with evals and controls built in from the first version.
Anatomy of a dSilo agent
Each building block is a governed, versioned object in Agent Studio. Business users configure it; engineers extend it; the control plane enforces it at runtime.
Defined skills, instructions, knowledge, APIs, MCP tools, business rules and actions — with explicit capability boundaries. Skills follow the open Agent Skills format, so a skill written once runs in dSilo apps, in Microsoft Copilot, or through the MCP Server. Tools come from the governed tool catalog with typed contracts, authorization and audit.
Deterministic workflows with bounded agent reasoning inside them: explicit handoffs between planner, specialist and validator agents, stop conditions, retries, timeouts and validations. Event-driven triggers start agents on a new invoice, a contract amendment or a price change — not only on a chat prompt.
Contracts, suppliers, customers, invoices, SKUs, obligations, pricing terms and transactions connected into one queryable business context. Retrieval combines enterprise RAG and GraphRAG over the procurement and commercial ontology, so every step of an agent sees the right slice of the System of Context, with citations.
Task context, durable checkpoints and execution state for long-running workflows that survive restarts and waits for approval. Approved organizational knowledge — categories, supplier history, negotiation outcomes, reviewer decisions — becomes long-term memory with scope and retention rules. Evidence is retained with every decision.
Validation, approval and escalation checkpoints inserted as first-class steps for contractual, financial, compliance and high-impact decisions — with owners, thresholds, SLAs and the evidence the reviewer needs. Checkpoints render in the Decision Hub, in Teams or in Copilot.
Continuous tests for grounding, tool selection, extraction, reasoning quality, workflow completion, policy compliance and outcome accuracy. Guardrails from the policy engine are enforced by the runtime, outside the model — the model proposes; the control plane decides what is permitted.
The agent lifecycle
Studio spans the whole lifecycle, so the agent that passes evaluation is the same object that runs in production — same identity, policies and observability.
Start from a business process — PPV recovery, renewal review, three-way match — and lay out the workflow: agents, tools, decision points and human checkpoints. Templates for Procure, Commercial and Operating Partner flows give a working baseline.
Bind agents to the Context Graph, select tools from the MCP catalog, attach knowledge and business rules, and set capability boundaries: what the agent may read, recommend, decide and execute.
Build golden sets from your own history, run eval suites on grounding, tool use, extraction and outcomes, red-team prompts and inputs, and set the pass thresholds that gate promotion. Every change re-runs the regression suite.
Assign agent identity and least-privilege access, set the autonomy level per scope, choose the surfaces — dSilo applications, Microsoft Copilot, MCP Server, custom experiences — and release under change control.
Trace every retrieval, model call, tool invocation, approval and outcome. Monitor cost, latency and drift. Reviewer overrides and measured outcomes flow back into evals and memory so agents improve safely.
Agent definition
An agent in Studio is a declarative definition, not a script. It captures skills, tools, boundaries, evals, autonomy and memory scope in one versioned artifact your team can diff, review, export and roll back.
# agent: ppv-recovery · scope: indirect / NA / MRO · v14 skills extract.invoice_lines · match.contract_price · draft.recovery_claim tools search_contracts · get_effective_terms · validate_invoice_pricing create_action (gated) · request_human_approval context graph: supplier, contract, po, invoice, payment · rag: pricing schedules workflow plan → extract → match → validate(agent) → gate(human) → act guardrails claim <= 50_000 USD → auto · else → queue:category_mgr evals golden=1_204 · grounding=0.99 · tool_select=0.98 · outcome=0.978 autonomy current=L2 · promoted 2026-06-14 · demote_if outcome < 0.95 memory checkpoints:durable · org_knowledge:approved · retention:7y · pii:masked surfaces procuregpt · copilot · mcp trace otel · retrieval, model calls, tool use, gates, overrides
Multi-agent collaboration
Studio composes agents into teams. Planner agents own the goal, specialist agents own a skill — extraction, matching, reconciliation, drafting — and validator agents check outputs before anything reaches a person or a system. Handoffs carry explicit contracts and evidence, so a long workflow finishes reliably instead of as one brittle prompt.
Own the task, sequence the workflow, and decide the next-best action within their capability boundary.
Extraction, matching, reconciliation, benchmarking and drafting — each a governed skill with its own evals.
Check grounding, policy and consistency on every output; route failures to retries or to a human.
A2A-style handoffs with typed contracts; event triggers from documents, transactions and system changes.
Semantic intelligence
Your systems of record store transactions and documents. Studio connects them into the context agents need — and keeps what the organisation learns.
Human-in-the-loop experiences
Configured, not coded. The same components render in ProcureGPT and CommercialGPT, in Microsoft Teams and Copilot, and as embedded pages in your own portals.
One queue for reviewing, approving, overriding and explaining agent recommendations — with the evidence and the clause or invoice cited.
Visual builders for reviewer assignment, escalation, loops, re-initiation and handoffs to other systems.
Alerts in Teams, email or Copilot when an agent needs a human decision, with the action available in place.
Agent activity, KPIs, autonomy levels and recommendations in real time, with click-to-filter and export.
Structured templates for analysis, plans and document review with side-by-side source evidence.
Auto-generated narrative and charts, every figure traceable to source records and agent runs.
Publish anywhere
An agent published from Studio is available as a dSilo experience, as an action inside Microsoft Copilot, and as a governed tool that any approved agent can discover through the dSilo MCP Server.
ProcureGPT · CommercialGPT · Decision Hub · dashboards and pages.
Ask, retrieve, analyse, act and continue in dSilo — without switching applications.
Typed tool contracts with OAuth / enterprise identity, tenancy, policies and audit for compatible agents.
Your assistants, enterprise apps and partner solutions through APIs and events.
Connected enterprise
Studio agents reach enterprise and external systems through MCP, APIs, events, databases, files and secure connectors — bringing context in and pushing approved actions back to the system of record.
Build with AI Native Services
A forward-deployed engineer configures your first agents and writes the evals, domain experts set thresholds and business rules, HITL reviewers validate context and data, and customer success runs training and adoption. Days to agents acting on live transactions.
Explore AI Native ServicesOpen by design
Agent definitions export as versioned artifacts. Skills follow the open Agent Skills format. Tools speak MCP. Handoffs follow A2A conventions. Traces emit OpenTelemetry. Models are routed across frontier and specialized providers by cost, latency and quality — your skills, evals and memory stay yours.
A two-hour working session on a sample of your data, or a scoped wave — first agents live in days — with an FDE, a domain expert and a CSM embedded with your team. Your choice.