Agents Decide. Workflows Execute.

How BitSafe Runs on Notion, Part 3 of 5

One of the easiest ways to make automation unreliable is to ask an AI agent to behave like a fixed workflow.

The reverse causes trouble too. A deterministic rule cannot resolve every ambiguous company name, interpret a meeting decision, or decide whether a piece of information belongs in a customer record.

We separate those jobs at BitSafe.

Agents handle interpretation. Workflows execute predictable steps. Notion holds the structured state both sides depend on, while NanoClaw provides bespoke computation, cross-source context, proactive work, and external execution when the task extends beyond the workspace.

This division matters more than the model choice. It gives each task a clear home and makes the overall system easier to test.

Before we automate something, we ask what kind of uncertainty it contains.

If the task requires reading context and making a bounded judgment, it may belong to an agent. Classifying a document, matching a meeting mention to an existing company, or drafting a summary from several sources all require interpretation.

If the steps and outcome are known in advance, a workflow is usually better. Setting a field after an approved state change, assigning an owner from a routing table, or moving data between two fixed schemas does not need open-ended reasoning.

Many useful systems combine both. An agent makes the decision, then a workflow carries out the mechanical steps. Keeping the boundary visible lets us review the judgment without pretending every action needs judgment.

Both agents and workflows need a dependable record of what is true.

Notion gives us typed properties, relations, page content, ownership, and status. An agent can interpret a messy input against that structure. A workflow can act on a precise state change once the input has been resolved.

Consider a meeting follow-up. The conversation may mention a company informally and describe several possible next steps. An agent can identify the relevant entity, summarize the decision, and propose the action. Once that proposal becomes a structured task with an owner and status, deterministic rules can handle reminders or routing.

The agent does not need to manage the entire lifecycle. The workflow does not need to understand the conversation.

That separation reduces the number of assumptions hidden inside either layer.

We get better results from agents when the scope is narrow.

A useful workspace agent might classify one field from approved context. Another might draft a document in a known format. A meeting agent might extract decisions and proposed actions without deciding which projects the company should pursue.

The key is to make the judgment bounded by a schema, an instruction source, or a clear approval point.

Broad mandates sound efficient because one agent can theoretically do more. They also make failures harder to diagnose. If an agent can interpret the request, change the data model, choose the recipients, and send the final communication, several different risks have been folded into one black box.

We prefer small decisions with explicit handoffs.

Once a process becomes stable, we remove reasoning where we can.

A routing table should route. A status transition should trigger the same next step each time. A scheduled check should evaluate a defined condition and surface an exception. The workflow can be tested against expected inputs without relying on a model to reproduce the same interpretation on every run.

This also controls cost and latency, though those are secondary benefits. The main benefit is operational clarity. When a predictable step fails, we can inspect the rule. When an agent makes the wrong judgment, we can inspect the context and instructions.

The two failure modes require different fixes.

Notion remains our system of record and collaborative workspace. Some work still needs more than workspace-native actions.

NanoClaw handles tasks that require bespoke computation, live context from several sources, proactive monitoring, or execution in external systems. It can use Notion as one source while combining it with tools that do not belong inside the workspace.

The distinction is complementary. A Notion agent is well placed to work with documents, database rows, meetings, and collaborative review. NanoClaw is better suited when the task needs code, broader source access, or proactive execution outside Notion.

We avoid describing either layer as universally autonomous. Human approval remains necessary for destructive operations and other high-risk actions. Automated customer-channel campaigns also follow a separate human approval workflow before distribution.

The goal is useful initiative inside clear boundaries.

A meeting follow-up shows the division of labor in a form most teams recognize.

First, the meeting record captures the source. An agent reads the summary and supporting notes, then identifies decisions, entities, and proposed next steps. Ambiguous matches remain visible for review rather than being silently forced.

Next, approved actions become structured tasks linked to the relevant project or company. Ownership and status now live in the shared operating model.

Finally, deterministic workflows handle the predictable follow-through. They can route the task, update a known field, or surface an overdue item according to fixed rules.

If the follow-up becomes a customer-channel campaign, the system can draft and assemble context, then routes the campaign through the marketing approval workflow before distribution.

Each layer does a smaller job. The chain is easier to understand because no single component claims the whole process.

We do not assume that a workflow should remain agentic forever.

A new process often starts with interpretation because the company is still learning the categories and exceptions. As the pattern stabilizes, more of it can move into rules, formulas, or fixed automation. The agent stays where judgment remains useful.

This is a practical maturity model:

The order prevents premature automation. It also stops the agent layer from becoming a permanent home for business logic that should be explicit.

When a workflow feels unreliable, we check whether it contains an unstated judgment. When an agent feels inconsistent, we check whether it has been asked to repeat a mechanical procedure.

Most problems become easier once the task is split at that boundary.

Agents decide within scope. Workflows execute known steps. Notion keeps the state coherent. NanoClaw reaches beyond the workspace when the task requires it. People retain control where consequences demand it.

That operating rule lets us expand automation without making the company harder to understand.

Subscribe to the BitSafe newsletter for more from the Canton ecosystem and occasional field notes on how BitSafe operates.