Introducing Sequence Labs
For the past several months, we've been scoping and building agents at Sequence. Today we're sharing some of that work publicly as Sequence Labs. This is where we share what we're prototyping, what's live, and what we've learned. Some of it will ship. Some of it won't.
How we build
LLMs are useful at different levels, and we think about each one differently.
The first is slotting them into specific product flows: a contract becomes structured data, a query gets a structured answer. The second is running guided automations with human oversight, so multi-step workflows with tools, review steps, and audit trails. The third is autonomous operation, where the guardrails are strong enough that the agent can run without someone watching every step.
We're building across all three. A few principles apply at every level, and they matter more the higher you climb.
The quality of your tools matters. Tools, context, and constraints are the biggest differentiators. A well-scoped agent with good context will outperform a general one with a better model.
Keep humans in the loop. Someone has to be able to see what the agent did, why it did it, and have a clear path to override it. In finance, this is the baseline.
Smallest useful scope first. Every agent should do one specific job well before we widen its remit. Most agent projects go wrong because the scope is too broad from the start.
Measure everything. "Vibes are good" isn't signal. We track quality, latency, and cost at every layer, on every agent we ship.
What's in Labs
Contract Intake Agent (Live, v0.37)
Customer contracts contain a lot of structured data: customer details, billing contacts, pricing terms, billing schedules. Until now, someone had to transcribe all of that into Sequence for every new customer.
Contract Intake Agent reads the contract and does it automatically. It extracts the key information and sets up the customer and billing schedule for review. It's live, it's in production, and it's been saving our customers real time since earlier this year.
Customer Knowledge Base (Prototype, v0.01)
Most of what a team knows about a customer lives scattered across Slack channels, shared docs, and call notes. This agent ingests those sources into a vector index and exposes it via an MCP server, so any agent can search and cite the source material rather than starting from scratch.
This one is still being built. We're publishing it because it's a building block for a lot of what comes next.
How we test ideas: Hack Day
In March 2026, we ran an internal hack day to form a view on how we want to build our next generation of agents. Going into it we had two questions to answer, and built seven prototypes.
The first question: how do we want to build our MCP server? This meant thinking through API key security and whether we could implement an OAuth IDP layer based on Stytch, scoped to RBAC rules per user.
The second: which agent framework do we want to adopt? We had multiple people try different TypeScript and Python frameworks, building agents that interact with our API via tool calls, to see how each one fits with our existing infrastructure. Having different people try different approaches was deliberate.
What we're not building
We're not building agents that do the billing. Sequence's deterministic billing engine, the part that handles pricing models, usage calculations, and the actual math, stays deterministic.
We're building agents that handle the work around billing: contract intake, customer queries, data retrieval. The jobs that are manual, slow, and error-prone. That frees up the finance team to focus on the decisions that actually require judgement.
Labs is where we figure out which agents are worth building, how to build them well, and what we're willing to ship. We'll keep adding to it.
Explore Sequence Labs: https://sequencehq.com/labs
Stay ahead of the curve.
See how Sequence can transform your billing.