Gemini 3.6 Pro
GPT-5.6 Sol
Claude Opus 4.801
A production agent system you don’t have to build and maintain.
- The agentic harness runs and improves continuously.
- No long-term infrastructure roadmap of your own.
- Cost and time back, from the first agent.
01 — MANAGED AGENTS · SDK + HARNESS + RUNTIME
The managed harness to build and run AI Agents in production.
You define the agent’s behavior and business logic. Cominty runs your agent code and configurable harness on managed workers—with persistent state, sandboxed tools, policy enforcement, and streaming events.
call_agent.pypython
from cominty_sdk import AsyncCominty# Auth: your API key + user ID.async with AsyncCominty() as client: run = await client.chat.start( agent_id="agent::support-assistant", message=ticket.body, ) print(await run.text())Managed runtime
Ticket #4821Escalated · Billing
“I was charged twice for my March invoice.”
Draft reply ready: confirms the duplicate charge, references invoice #INV-3391, proposes a refund within 3–5 business days.
agent::support-assistant · 1.2s · $0.014
Build the agent from the dashboard below. Calling it is four lines of Python.
Trusted by teams running production AI agents on Cominty
- Entrepreneur Invest
- Groupe ADSN
- Groupe Actual
- OMS
- Spinpart
- Eight Advisory
- Berger Levrault
- Axys
- Values Media
- IAC Partners
- Zcomme
Create an AI agent in minutes.
Five steps, the real product flow — identity, instructions, models, context, review. Select any step to replay it. You’re building on the Cominty harness, not maintaining a homegrown loop — it arrives already tuned for whichever model you pick, so swapping models later doesn’t mean rebuilding your stack.
Agents — create one, no codeLive
HR Scoring Agent is live
In the composer’s picker for everyone with access.
Agent ID · use it in every call
agent::hr-scoring-agentCall it from the SDK
call_agent.pypython
async with AsyncCominty() as client: run = await client.chat.start( agent_id="agent::hr-scoring-agent", message=request.text, ) print(await run.text())Done · no code writtenAgent created
Yours to call
HR Scoring Agent is live. Call it by its ID from the SDK — the harness, the state and the permissioning are already running underneath it, tuned for the model you picked.
- HR Scoring Agent — 2 tools · 1 sources — Claude Opus 4.8
- Finance Agent — 3 tools · 2 sources — Gemini 3.6 Pro
- Operations Agent — 4 tools · 3 sources — GPT-5.6 Sol
- Legal Review Agent — 2 tools · 4 sources — Claude Opus 4.8
- Sales Ops Agent — 3 tools · 2 sources — Mistral Large 3
- IT Support Agent — 5 tools · 2 sources — DeepSeek V4
Build the AI agent once.Call it from anywhere.
Every use case after the first one is a few lines of Python — the same call with a single thing changed.
One agent· agent::finance-analyst
weekly_digest.pypython
async with AsyncCominty() as client: run = await client.chat.start( agent_id="agent::finance-analyst", message="Revenue by region, last week.", disabled_tools=["web"], # internal only ) digest = await run.text()slack.post("#revenue", digest) # or Power BIWhere the call lands
Finance analyst · Monday 09:00
Week 27 revenue · +6.2% week on week
- EMEA+11%
- North Americaflat
- APAC−2%
Support assistant · thread memory
3 threads open · 41 turns held
- Longest thread18 turns
- Context carriedlast 6 turns
- Session store you operatenone
Skill · v4 · owned by Finance
4 steps · 2 approval gates
- Reconcile bank feeddone
- Match invoicesdone
- Controller sign-offwaiting
Streaming · 6 events
search_docs → read_invoice → draft_reply
- search_docs214ms
- read_invoice380ms
- draft_reply1.1s
weekly_digest.pyA cron job, an agent call and a webhook: the weekly digest lands where the team already works, with no interface to build.
The result lands wherever your business already runs: Slack, your BI, your ERP, a webhook, a feature of your SaaS, an in-app copilot, an API endpoint, a background worker in your backend.
Runs that survive whatever broke.
A run is not a fire-and-forget call. It is a durable session your team can inspect, resume and stop: state survives crashes, timeouts and pauses, and a human stays in the loop where it matters.
resume_run.pypython
async with AsyncCominty() as client: run = await client.chat.get(run_id="run_8f2c…") if run.status == "waiting_approval": # a human reviews the pending call await run.approve(step_id=run.pending_step) elif run.status == "failed": # state is intact — resume the run await run.resume()›run.status→"waiting_approval"
Run timelinerun_8f2c… · 5 steps
- Fetch invoices312msState intact
- Validate totals88msState intact
- Calculate refund1.4sState intact
- Post to NetSuiteattempt 2State intactRetried
- Send confirmationapprovalState intactWaitingApprove
Waiting on human approvalView full trace & replay
The prompt is the easy part.
Production infrastructure is where most agent projects stall. None of it shows up in a demo, and all of it has to exist before an agent can run unattended.
Without Cominty
Building it yourself
A slice of the stack — and it keeps growing.
- Hosting & scaling
- Session state
- Sandboxing
- Credentials
- MCP connectors
- Security & compliance
- Cost & usage monitoring
- Multi-LLM routing
- Observability
- Memory management
- Evaluation & guardrails
- Audit logging
- Incident response
- Access governance
- Loop system
- File & vector storage
- Rate limiting & retries
- Secrets rotation
- Queue & scheduling
- Compliance reporting
- …and more, every release
Yours to build, run, patch and staff — before the first agent ships.
Absorbed
With ComintyLive
Running on Cominty
Managed agent execution layer
deploy.pypython
agent = cominty.Agent( task="Reconcile invoices", tools=["stripe", "netsuite"],)agent.deploy() # live in prodYour team builds the agent. Cominty builds and runs everything underneath it.
Everything a security team asks for in the first call.
Cominty is built the way a CISO would build it: isolated by default, encrypted everywhere, infrastructure-agnostic, and never trained on your data.
Audited annually
SOC 2 Type II
ISO 27001 2022
GDPR Compliant
EU AI Act Compliant
- GDPR compliant — Data residency and processing controls native to every deployment. (DPA & sub-processor list)
- SOC 2 Type II — Independently audited controls across security, availability and confidentiality. (SOC 2 Type II report)
- ISO 27001:2022 — Information security management system aligned with the 2022 revision of the standard.
- EU AI Act compliant — Commitments, measures and processes mapped to Regulation (EU) 2024/1689, with ethics by design. (AI Act compliance dossier)
- Data & residency
Encrypted
At rest: server-side AES-256 on every stored artifact and secret. In transit: TLS 1.2+ enforced on every external connection.
- Data & residency
Never trains on your data
No training, no fine-tuning — ours or our providers'. API-only, contractual no-training terms, zero data retention at OpenAI.
- Data & residency
EU-hosted by default
AWS EU regions with data residency guarantees; every sub-processor listed in the Trust Center.
- Isolation & operations
Full audit trail
Every agent action, tool call and approval is logged with who, what and when — exportable to your SIEM.
- Isolation & operations
Tenant isolation
Strict per-organization data scoping, least-privilege tool access.
- Isolation & operations
Tested & monitored
Third-party penetration tests at least annually, intrusion detection on the production network, MFA on every remote access.
Want the full documentation — pen tests, audit reports, controls status…? Everything a procurement or security review needs, in one place.
Go to Trust Center (opens in a new tab)Why IT teams choose Cominty
IT and product teams want to accelerate AI adoption without disrupting their existing business roadmap or compromising their cost and security requirements. Cominty gives them the simplicity and flexibility to integrate AI quickly, without adding another layer of complexity.
- 02
+120 models. One API.
- One API, one contract, one bill.
- Switch model without a rewrite — no migration cost.
- Automatic failover between providers.
- Bring your own keys.
- 03
Full cost transparency.
- Pay for execution: EUR 0.30 per run.
- The rate falls as volume rises.
- 0% markup on model tokens.
- Cost dashboard and per-run monitoring.
- 04
Context, plugged in.
- Connectors and any MCP server.
- Register your own skills.
- A RAG pipeline for unstructured documents.
- Context added to an agent in one line.
- 05
Memory, managed.
- Memory scopes you define: per agent, per user, shared across a team of agents.
- Read and written on every run, governed like everything else.
Ask us anything.
Transparency is at the heart of our solution. We’re ready to answer any questions you may have.
You can. The question is what it costs you. Durable state and retries, sandboxed execution, approval gates, tenant isolation, per-step tracing, evaluation and regression checks, model failover across providers — that is the layer teams spend months on before their first agent reaches production, and it never stops needing maintenance. None of it differentiates your product. Cominty ships it built and governed, so your engineers work on the part that is actually yours.
Ship your first agent
From prototype to production in days, not months.
Configure it from the dashboard or in four lines of Python. Cominty runs everything underneath it.
- No infrastructure to stand up
- Any model, no lock-in
- SOC 2 Type II · GDPR