Plan validation
Plans are checked before they run: risky steps without an approval gate, bad tool restrictions or missing undo plans fail at plan time, not mid-mission.
Insiro Agent Harness
The harness that turns models into dependable agents.
Describe a job in plain language. Agent Harness builds the agent, picks the right model for each step, gives it tools, a browser and a sandbox, and runs it behind permissions, budgets and an audit trail.
OpenAIAnthropic ClaudeGoogle GeminiInsiro modelsBYOK
What is an agent harness?
An agent harness is the runtime around a model that turns it into a dependable worker: it supplies tools, memory and knowledge, routes each request, enforces permissions and budgets, verifies results and records what happened. Insiro Agent Harness provides all of that so teams build agents instead of infrastructure.
- 13governance checks on every tool call
- 4risk tiers with approvals
- BYOKOpenAI, Claude, Gemini or Insiro models
- 1audit trail for every step
Agent builder
Write instructions. Get a working agent.
Explain the task the way you would to a new hire. Insiro drafts the agent: steps, tools, knowledge sources, model and schedule, ready to test in a side-by-side preview.
- No-code builder with live preview, or import from OpenAI Assistants, Claude, Gemini or JSON.
- Knowledge and memory: attach documents, sites and cloud files for RAG, plus personal and org memory with governance.
- Publish to the Agent Hub, to a team Space or to Microsoft Teams, with visibility you control.
Why Insiro, not the alternativesMost builders stop at a prompt and a tool list. Insiro agents ship with a model route, a budget, a risk policy and an audit trail from the first run, so the agent you demo is the agent you can deploy.
1. People profile search
Get the requesting employee's profile
2. Branch
Agent decides which path to follow
New employee
3. Respond
Explain they are not yet eligible for this review cycle
Fallback
4. Plan & execute
Let the agent choose actions to answer
SharePointRead files
Read only
SalesforceSearch accounts
Read only
GmailDraft replies
Needs approval
SlackPost summaries
Needs approvalEach tool runs under the user's own login, through the governed gateway.
Is Acme eligible for the Q4 renewal discount?
Yes. Acme renewed early and is above the 50-seat threshold, so the 8% discount applies. Source: Renewal policy v3 (SharePoint).
fast model2 tool calls$0.004
Agent HubDiscoverable by your org
Sales SpaceOnly the sales team can run it
Microsoft TeamsChat with it where people work
ScheduleWeekdays at 8:00
Auto routing
The right model, the right tools, for every request.
A front-door router classifies each request into one of eight intents, weighs reasoning, quality, modality, complexity, cost and task type, then picks the model. Click a request, or change the model controls.
Insiro Routerintent · cost · quality
Auto routing
Reasoning levelQualityModalityComplexityCostTask type
Model controls
Claude Opus 5Claude Sonnet 5
Gemini 3.5 Flash
GPT-5.6 SolInsiro FastClarification gate
Below 70% confidence or with a missing entity, the agent asks ("Who should I send this to?") instead of guessing.
Task decomposition
Multi-step requests split into ordered sub-tasks, and apps you still need to connect are flagged up front.
Continual learning
Tool ranking gets a small, capped nudge from verified outcomes in your organisation.
Why Insiro, not the alternativesSingle-model assistants pay reasoning-model prices for simple lookups and still guess on ambiguous asks. Insiro routes by intent and cost, and stops to ask rather than act on a wrong assumption.
Company context
Answers grounded in who knows what.
Agents work from a personal graph and an enterprise graph: people, projects, customers, data and documents, linked by how they actually relate. Hover a name in the answer to see the connection behind it.
Is anyone already building a way to predict which customers will churn next quarter? Who should I talk to first?
Show work · 3 sources
Yes. Two efforts overlap and don't seem to know about each other. Start with:
- Priya Raman, the only person contributing to both Churn Signals and the Customer Health Score.
- Marco Lindqvist, who ran the shelved Risk Alerts Pilot and wrote the first spec.
- Both read from usage_daily and Zendesk tickets; Revenue Ops owns the score today.
PersonProjectDataDocumentTeam
Personal and org memory, governed by admins RAG filtered by each person's access Datasight semantic layer shared by every agent
Mission runtime
Everything a long-running agent needs, built in.
Agent Harness runs every multi-step job as a mission: an objective, acceptance criteria, a cost ceiling, a versioned plan and verified progress.
Capability tokens
Each task attempt gets a short-lived, revocable token bound to its user, mission and tools. Child agents can never get wider access than their parent.
Verification ladder
Each task's output is verified before downstream tasks use it, and the mission produces an acceptance report against your criteria.
Compensation (undo)
Side-effecting steps declare how to reverse them. If a mission fails mid-way, Insiro can roll back what it changed.
Replanning
When reality changes, the agent proposes a typed patch to its plan. Every plan version is kept with the reason it changed.
Idempotent actions
Server-derived intent keys mean a retried step never sends the same email or creates the same ticket twice.
Why Insiro, not the alternativesAgent frameworks give you primitives; you still build validation, retries, idempotency, rollback and audit yourself. In Insiro these are the default runtime, tested across every agent, teammate and pipeline on the platform.
Skills
Teach it once. Your whole team reuses it.
Turn on Teach mode and do the task yourself in Insiro's cloud browser. Insiro records the steps, the apps and what to avoid, and saves it as a skill agents can run on demand.
- Teach by demonstration: no flowcharts to draw, no selectors to maintain.
- Share like a document: restrict to people and groups, or open to the organisation, with auto-enable.
- Guardrails travel with the skill, so every run follows the same rules.
Why Insiro, not the alternativesPrompt libraries drift and RPA scripts break when a page changes. An Insiro skill is a tested procedure that agents adapt at run time, with the same permissions every time.
skills / meeting-prep / SKILL.md2.1k runs
Meeting prep
Brief on attendees, history and talking points
Shared with Sales teamskills / deal-desk-check / SKILL.md640 runs
Deal desk check
Validate discount and terms against policy
Shared with Sales teamskills / quarter-close-checklist / SKILL.md312 runs
Quarter-close checklist
Run the 18-step close and flag blockers
Shared with Finance teamskills / invoice-triage / SKILL.md1.4k runs
Invoice triage
Match bills to POs, route exceptions
Shared with Finance teamskills / pr-reviewer / SKILL.md3.8k runs
PR reviewer
Review for tests, security and style
Shared with Engineering teamskills / incident-commander / SKILL.md210 runs
Incident commander
Open the war room and timeline
Shared with Engineering teamskills / refund-triage / SKILL.md980 runs
Refund triage
Investigate claims against policy
Shared with Support teamskills / brand-voice-reply / SKILL.md5.2k runs
Brand voice reply
Draft replies in your tone
Shared with Support teamExample skills. Teach your own by demonstration, then share them like a document.
Cloud browser and sandbox
Heavy work runs inside your walls. Only the answer comes out.
When a site has no API, the agent opens its own cloud browser that you can watch live and take over at any point. When a task means thousands of rows, it writes and runs code in an isolated sandbox.
- Live view and takeover: watch at up to 24 fps, take control, then hand back.
- Safe by default: the agent browser cannot submit or post without approval, cannot reach private addresses, and never types into password or card fields.
- Coding agent with a full dev workspace, package allowlist, typosquatting and vulnerability scanning.
Your organisation
Snowflake SharePoint NetSuite
Snowflake SharePoint NetSuite Private sandboxcode runs here · 4,812 rows
Only the answer leaves17 exceptions · 1 file
Egress to the open internet blockedallowlist: 0 hosts
Reconcile October's Stripe payouts with NetSuite
Sandbox · reconcile-octoberisolated · no egress
Download payoutsMatch the ledgerFlag the exceptions
Payouts checked0 / 4,812
https://portal.vendor.example/invoices
Invoice INV-2207 · $4,120
Invoice INV-2208 · $980
Invoice INV-2211 · $12,400
Blocked: the agent tried to submit a payment form. Submits need your approval.
Password and card fields are never typed into by an agent.
Take over3 requests blocked so far
Governed tool calls
Every tool call takes the same 13 steps.
Agents never call tools directly. They ask the Insiro gateway, which checks the request, brokers the user's own credentials, reserves cost and writes evidence, in this order, every time.
Why Insiro, not the alternativesHomegrown MCP proxies usually cover auth and logging. Insiro adds task-scoped grants, DLP on input and output, idempotent replay, cost reservation and fail-closed audit, so a missing log line blocks the call instead of hiding it.
No tool runs without consent
- Registry lookup
- Capability grant
- Space entitlement
- Kill switch
- Rate limit
- Schema check
- Input DLP
- Credential broker
- Risk tier & approval
- Idempotent replay
- Cost reservation
- Output DLP
- Audit evidence
Why Insiro
Agent Harness vs. the alternatives.
You can assemble an agent stack from a framework, a vector database, an observability tool and a proxy. Or you can start with one that is already governed.
| Capability | Insiro Agent Harness | DIY agent framework | Copilot agent builders |
|---|---|---|---|
| Time to first governed agent | Same day, no code | Weeks of platform work | Fast, limited actions |
| Multi-model routing with BYOK | Intent and cost based | You write the router | Single vendor model |
| Risk tiers and exact-action approvals | Built in | Build it | Basic confirmations |
| Task-scoped capability tokens | Bound, revocable, cascading | Build it | Not exposed |
| Cloud browser with human takeover | Live view, takeover, safe defaults | Add a browser service | Rare |
| Isolated code sandbox with supply-chain checks | Egress allowlist, package policy | Add a sandbox vendor | Limited |
| Rollback of side effects | Compensation plans | Build it | No |
| Hash-chained audit evidence | Fail-closed | Build it | Activity logs |
What is an agent harness?
An agent harness is the runtime around a model that turns it into a dependable worker: it supplies tools, memory and knowledge, routes each request, enforces permissions and budgets, verifies results and records what happened. Insiro Agent Harness provides all of that so teams build agents instead of infrastructure.
Which models does Insiro Agent Harness support?
Insiro routes across Insiro-hosted models and models from OpenAI, Anthropic Claude and Google Gemini. Organisations can bring their own API keys (BYOK), set a default model per agent or teammate, and let the router pick a fast, reasoning or coding model per request based on intent and cost.
Can I import agents I already built elsewhere?
Yes. Agent Harness imports agents from OpenAI Assistants, Claude and Gemini configurations or a JSON definition, keeping their instructions and mapping their tools to Insiro connectors. Imported agents immediately run under the same governed gateway, budgets and audit trail as native agents.
How does Insiro stop an agent from doing something risky?
Every planned action is classified into four risk tiers. Tier 0 and 1 run immediately; tier 2 needs user confirmation; tier 3 needs owner approval and the user's own credentials. Plans containing risky steps are rejected unless they include an approval gate, and each agent carries a task-scoped capability token that limits tools and arguments.
What are skills in Insiro?
Skills are reusable procedures taught by demonstration. You perform a task once in Insiro's cloud browser, Insiro records the steps, apps and guardrails, and agents can then run that skill on request or automatically when relevant. Skills can be shared with people, groups or the whole organisation.
Can agents write and run code?
Yes. Agents run code in an isolated sandbox with an egress allowlist, package policy and vulnerability scanning, and a coding agent gets a full development workspace per run. Results come back as files, tables or answers; destructive commands such as rm -rf or terraform destroy are classed tier 3 and need approval.
Build your first governed agent this week.
A 30-minute working session: we connect one data source and one app, then run a governed agent on a task you choose.
- 13governance checks on every agent tool call
- 1,700+SaaS apps agents can act in
- 30+data dialects for plain-English questions