From answering questions to running the work

The first release, on January 21, put a governed front door in front of the warehouse. Any assistant that speaks MCP could connect. People signed in with the company’s own identity provider, roles decided what each person’s assistant could reach, and every call was recorded. What made it more than a database driver was the context: each answer arrived with the meaning, ownership, and warnings attached to the columns it used, so the assistant did not have to guess what a field meant.

This was Deasil’s long-standing view of company data, a thesis from an internal project called Methodology, rebuilt on an open protocol that every frontier assistant was adopting. The history page tells that longer story.

The second stage, February through June, made the context layer learn. In February an assistant could capture what it learned about the data and a person could review it before it became shared. In March the portal arrived, so the reports and dashboards agents produced had a home people could open, version, and share, and the platform reached 1.0 on March 9. Memory across sessions followed in April.

Reach widened beyond the warehouse. In April the platform began routing to other MCP servers through one governed gateway, and in May it learned to reach any REST API from its specification. June brought the prompt library, so the procedures a team relied on became versioned and shared, and knowledge pages, where the business facts that belong to no single table live. By the end of June, what one person taught the platform reached every person’s assistant.

The third stage moved settled work out of the conversation. From August 17 the agent could write automations that the platform stores, versions, and runs without a model in the loop. Spreadsheets became tables a week later. September added presentations, notification channels, events arriving from other systems, and the rule that matters most for trusting an agent’s code: when a new version would behave differently, a person approves the change from a plain-language summary. Signed delivery of results to other systems arrived on October 6.

This is where Plexara started doing work that companies usually buy separate tools for, a reporting subscription here, a sync service there. It did not replace the systems the work touched. It became the place the work between them ran.

The fourth stage, the second half of October, filled in what those automations needed to live on their own. Slack joined the channels, and street maps began to be served from each deployment. A failing automation started sending one alert per failing streak instead of one per run. On October 14 Plexara’s own catalog of every connection arrived in the portal, with lineage drawn by the work that runs through the platform and a history of every change.

Then the platform learned to receive. Automations could answer requests at their own address, read a mailbox, run once for every event a source stores, and power forms the agent builds. Portal permissions per role and access requests arrived, and on October 25 people outside the company could sign in through the company’s identity provider to see exactly what was shared with them, with no access to any agent. Each of these does a job a company would otherwise buy a separate tool for, and each is used one workload at a time.

All of this shipped as 377 releases between January 21 and October 29, 148 of them feature releases, about forty a month. The count understates the later months, where a single release often carried several features. The number matters less than what it says about operations: two client deployments went live in March, around the 1.0 release, and have taken every release since. In late October all five deployments moved to the newest release within a day of each other. A managed service is what makes that cadence safe for the customer. Deasil tests, rolls out, and monitors every release, and nobody on the client side schedules an upgrade.

Through all four stages a few things did not change. The protocol is open, so the assistant in front is whichever frontier model a team prefers, Claude, ChatGPT, Gemini, and the ones that replace them. Access starts closed and is granted on purpose. Every call records who made it and why. And knowledge is taught once and reaches everyone, which is the idea the whole platform grew around.

The changelog carries the story forward two weeks at a time. For where the platform sits now, see the AI-first data platform.