Haltio pauses high-risk agent actions — refunds, deletes, sends, deploys — for human approval before they ship. Open source. 2-minute integration.
Hi I'm Devesh, I spent my life savings building an ecommerce store "Kawaii404.com". To scale it, I invested heavily in AI agents — automating support, refunds, and order handling, as I was confident in AI as a Computer Science Student.
One day, I discovered the agent had refunded 80% of the store's profit back to customers, triggered by a single misread policy edge case. Nobody had approved it. Nobody even saw it happen until the damage was done.
Agents should move fast. But the actions that can actually hurt a business need a human checkpoint first.
Agents executing irreversible actions with no review step before they happen.
Too many low-value interruptions train humans to rubber-stamp everything.
Pilots stall because teams don't trust agents enough to actually deploy them.
When 20 agents hit a decision point at once, who decides first — and what depends on what?
"Enterprise AI pilots are not failing because agents are incapable. They are failing because teams don't trust them enough to deploy."
A risk classifier sits between your agent's decision and its execution. Operators (the developers or compliance leads who configure agent permissions) define what's safe to automate — and what isn't.
LOW-risk tasks proceed automatically only if they fall within the agent's configured permission scope, defined by the operator at setup.
Timeout, escalation path, and fallback behavior are configurable per agent and per task type. Example: timeout_seconds=300, on_timeout="escalate"
When 50 agents hit a checkpoint simultaneously, Haltio queues, prioritizes, and routes approvals without blocking your pipeline. Dependency-aware. Timeout-safe. Rollback-ready.
Haltio isn't a single-agent off switch. It manages approval dependency graphs across many concurrent agents at once — queuing, prioritizing, and routing decisions in real time so one bottleneck doesn't stall your whole pipeline.
High-stakes checkpoints surface first when many agents request approval at once.
Tracks which agent actions depend on which approvals, so nothing executes out of order.
Set a timeout per task type, with a defined fallback if no human responds in time.
One agent waiting on a checkpoint doesn't freeze the others.
Choose what happens if Haltio itself goes down — block everything, or proceed with full logging.
Approve or block directly from Slack — no context-switching to a separate tool.
One decorator. Any agent framework. No architecture changes.
from haltio import checkpoint
@checkpoint(risk="high", on_timeout="escalate", timeout_seconds=300)
def process_refund(order_id, amount):
...
Haltio wraps any agent's action-execution layer with a single decorator. No framework lock-in, no architecture changes.
Checkpoint payloads are encrypted in transit (TLS 1.3) and at rest.
No task data is stored by Haltio beyond a configurable retention window.
Yes — self-hosted deployment is available for data-residency requirements.
Configurable fail-safe (block all pending actions) or fail-open (proceed with full logging). Enterprise tier targets 99.9% SLA with multi-region redundancy.
You can — for a single agent on a single framework. LangGraph's interrupt() and the OpenAI Agents SDK's native approval flow both work well there. Haltio is for the moment you have agents built on more than one framework, or more than one agent hitting checkpoints at the same time — that's a vendor-neutral, cross-framework problem no single framework owner is incentivized to solve for its competitors' agents too.
What you see here is a working, production-ready MVP — a real open-source SDK, not a mockup. It's Phase 1 of a much larger system: a full multi-agent orchestration and enterprise compliance platform. The roadmap below shows where this is going. We're looking for developers to help build it.
Three phases, in order, with no skipped steps.
Open-source Python SDK with human-in-the-loop checkpoints.
Concurrent approval graphs, SLA management, escalation logic.
RBAC, SSO, audit trails, EU AI Act reporting.
The EU AI Act mandates human oversight for high-risk AI systems. Enterprise security teams increasingly require audit trails before approving any agentic deployment. This is a tailwind, not a headwind, for a product like Haltio.
Figures vary by report scope. Research and Markets projects the AI governance and compliance tooling market at roughly $0.6B in 2026, growing to roughly $2.6B by 2030 at a ~44% CAGR. Other analyst firms scope the market differently and land higher — we're citing the conservative, named estimate rather than averaging conflicting reports into a number nobody can trace.
AI agents that touch money — refunds, payouts, order changes — inside teams shipping their first production agent.
Series A–C B2B SaaS companies with a dedicated ML/AI engineering team (5–50 engineers), deploying LLM-based agents into production workflows for the first time, where a single agent mistake has direct customer or financial impact.
Draft wedge, not yet confirmed against the beta cohort — narrow this to wherever the 5+ beta developers are actually concentrated (e.g. a specific vertical) once that's known.
Reserved for quantified testimonials:
Like Sentry, we start with a free open-source SDK developers add in minutes. The enterprise layer — audit trails, RBAC, compliance dashboards — is where revenue lives. The SDK is the top of funnel; the platform is the business.
| Tool | What it actually is |
|---|---|
| LangGraph interrupt() | Single-agent, synchronous only, no dashboard |
| OpenAI Agents SDK | Native pause/resume approval flow, but scoped to OpenAI's own SDK — no cross-framework or cross-vendor layer |
| HumanLayer | Framework-agnostic approval routing over Slack/email — single-agent, no concurrent dependency graph |
| Prefect / Temporal | Workflow orchestration, not agent-specific, no Slack-native UX |
| Build-it-yourself | Months of engineering, no standard protocol |
| Haltio | Multi-agent concurrent, SDK + dashboard, 2-minute integration |
We're not first to ship human-in-the-loop approvals — LangGraph, the OpenAI Agents SDK, and HumanLayer all support it today, and we'd rather say that plainly than have you find it yourself. What none of them do is manage approval dependencies across many concurrent agents running on different frameworks in the same pipeline. That's the gap Haltio fills, and why a framework or model vendor isn't the one positioned to build it — a vendor-neutral audit layer has to work the same way across competitors' agents too, which isn't something any single framework owner is incentivized to ship.
approvals.json plus a webhook is a weekend build for a single agent — we know, because that's roughly what the simplest version looks like. What's harder to replicate is the layer above it: a dependency graph tracking which actions depend on which approvals across many concurrent agents, configurable timeout and escalation logic per task type, and — on the roadmap — a policy engine with fine-grained, per-action permissioning that makes Haltio the one integration point across multiple agent frameworks inside an org. That's sustained orchestration and reliability engineering, not a weekend script.
Unlimited agents, community support.
Per-workspace monthly fee. Team dashboard, Slack integration, audit logs.
Custom pricing. RBAC, SSO, compliance reporting, SLA.
Engineering — build the production multi-agent orchestration backend.
Developer growth — open-source community, docs, integrations.
First enterprise pilot — dedicated customer success.
Raise amount to be finalized directly with Devesh — we'd rather have that conversation by email than publish a placeholder number on this page.
Reserved for the milestone story tied to the ask: e.g. "Raising $X gets us to Y design partners, Z paying enterprise pilots, and a shipped policy engine (Phase 2) within 12 months." Partners are funding what the raise buys in a year, not the product as it exists today — this is where that gets stated explicitly.
We have a working SDK, 5+ developers already testing it, and a team that has lived this exact problem firsthand. If you invest in developer infrastructure, let's talk — email devesh@haltio.in.
Building governance infrastructure for AI agents after watching firsthand what happens when nobody's watching: his uncle's ecommerce store lost 80% of its profit to a single unapproved agent action. That's not a hypothetical case study — it happened to family, which is why Devesh is building the fix now, while agent adoption is outpacing the governance layer underneath it, instead of waiting for someone else to.
LinkedIn - Click here to connect with meCommitment level isn't stated yet. "Full-time since [month/year]" reads very differently from a side project to a partner deciding whether to back the person, not just the idea — fill this in directly rather than leaving it implied.
Developers currently in active beta.
Reserved for one named design partner: company name, logo, and one quote attributed to a real person Devesh can stand behind. A single named company — even a 10-person startup, even unpaid — carries more weight here than any amount of polished copy.