The agent thatcan't break your rules
Sophisticated work is delegable because each of these is stopped by structure.
- Pay invoice: $8,120 (cap $5,000) — A general-purpose agent: Paid in full. No one was asked. On Perstack: Caught over the cap. Waiting for approval. (Mechanism: Approval checkpoint)
- Post invoice #4211, again (already posted) — A general-purpose agent: Booked twice. On Perstack: Blocked by rule. Nothing was posted. (Mechanism: Pre-execution rule check)
- “clean up old records” (production data) — A general-purpose agent: Bulk-deleted production data. On Perstack: Agents can only call operations you've defined. (Mechanism: Allowed operations only)
- call an external system (borrowed credentials) — A general-purpose agent: Reached systems it was never meant to touch. On Perstack: Every external call goes through the secure gateway. (Mechanism: Credential-free gateway)
- A bad operation corrupts data (09:41) — A general-purpose agent: What ran, what it touched, why: unknown. On Perstack: It's all on record. (Mechanism: Tamper-proof audit record)
Why Perstack
Rules it can't break are why you can delegate
What you delegate to an agent today is limited by what a person can supervise. A Perstack agent is structurally incapable of breaking your rules, and every run leaves a record that proves it didn't. Sophisticated work becomes work you can delegate. Once delegation is safe, throughput stops being a function of headcount.
Showcase
AI-native, in practice
What got automated, and which rules held. Four cuts through the work running on Perstack today.
Technology
Worth delegating to
Safe is the floor. What makes an agent worth the handover is the runtime under it.
Governed Agents
Every run gets its own throwaway machine, and the credentials stay outside the agent. An agent that goes wrong takes nothing with it — which is what makes it reasonable to give one a real task instead of a demo.
Agentic Backend
Your records sit in a database no other tenant touches, and work that runs for days survives restarts because progress is checkpointed. That is what lets you hand over a process, not just a task.
Platform
Handed over, and still running
The machinery that keeps handed-over work running is all here.
Capability
Context management
Runs are checkpointed step by step and can resume from where they stopped. Long-term memory accumulates what the work has learned.
Personalization
Runs and data can be scoped to each end user. Every record keeps who the run was acting for.
MCP & skills
Agent abilities extend as skills. Through MCP connections, internal tools and external services become instruments the agent can use.
Continuity
Event-driven execution
Cron schedules, API-triggered starts, resumes on approval. Every event that should move the business turns into a run, without being dropped.
Webhook delivery
Business events are delivered to external systems as signed webhooks, with retries and dead-letter management built in.
Audit
Every run lands in an append-only record. Tampering is rejected at the database level, and you can audit the full population.
Getting started
Describe the work. Connect the systems.
You describe the work to an agent in Perstack and it builds the system. Then you point it at the services the work already runs on.
Build it by describing it
You
Handle inbound supplier invoices: match each one to its purchase order, flag mismatches, and hold anything over $5,000 for approval.
Perstack
Built it. Invoice and PurchaseOrder records, a mismatch check before posting, and an approval step above $5,000.
What came back
- 6 records
- 4 operations
- 1 approval step
- 2 screens
Say what to change and it changes. Every version stays on the record.
Connect what it runs on
Declare the endpoints once. Credentials stay on the platform and never reach the agent.
Gmail
Read inbound mail, send the reply
Slack
Notify the approver, take the answer
Your API
Any HTTP service you already run
Gateway
Credentials held here
Agent