WorkSwarm 0.2.7: AI Agent Teams Hit Context Limits

On September 30, 2026, the openJiuwen community released WorkSwarm v0.2.7.

It is an open framework in which multiple AI agents work as one team. Beta 0.2.8 followed soon after.

The release's main changes are Proactive Context Service and sessions shared by multiple agents.

The authors invested in how agents share knowledge, and that is the right bet. In our projects, multi-agent pilots primarily stall when agents lose context while handing off a task; the model usually performs well.

In the usual setup, the agent retrieves data when it realizes they are needed: it searches the knowledge base (RAG) or calls a tool.

If the agent does not think to ask, it builds an answer without the required data. PCS reverses the order.

According to the release description, a separate service gathers task data in advance and passes it to the agent before the request.

The release notes contain no detailed PCS architecture, so the quality of data selection remains a hypothesis that must be tested on your own tasks. Cross-Agent Sessions handle task handoffs between agents.

The next agent sees the previous agents' history and interim conclusions and continues from where they left off. SkillPack is a packaged agent skill: an instruction plus a set of tools reused across different teams.

Self-evolution means that the system modifies its own skills based on completed tasks.

For production, this is the riskiest mechanism.

If a skill changes without regression tests, the agent behaves differently a month later than it did on acceptance day, and no one can identify why.

WorkSwarm handles orchestration.

Three questions remain with the company: 1. Data access.

Proactive context is only as useful as the completeness of its sources.

If stock is in 1C, product attributes are in PIM, and orders are in Magento, someone must connect these systems with the correct permissions. 2. Security.

A trial model is suitable for getting acquainted, but operational data must not be sent through it.

A CIS company handling customers' personal data under GDPR needs control over which fields are sent to an external LLM. 3. Governability.

Skill self-evolution requires versioning, testing, and rollback.

Without them, the CIO cannot be accountable for the system's behavior.

At KT.Team, we build agent systems the same way we have built integrations in recent years: the data bus and access rules first, then the smart components. System access we provide to agents through MCP, a standard protocol that lets the model call tools with typed parameters. The agent receives the operation "find counterparty" in 1

Or “get SKU attributes”

in Akeneo and Pimcore, without direct database access.

If the company already uses Apache Kafka, Datareon, or MuleSoft, we turn bus events into that proactive context: the agent receives an order change immediately, without polling. Security is handled by the LLM & Security Gateway.

The gateway masks personal data before sending it to the model, logs requests, sets budget limits, and assigns tasks among GigaChat, YandexGPT, Qwen, and Claude based on data requirements and cost. Shared context is clearly visible in PIM projects.

Generating item card text, tagging the taxonomy, and configuring catalog facets are three different tasks, and all three rely on one product model. When agents read it from a single

WorkSwarm 0.2.7 is worth installing in a sandbox to test how PCS selects context for your tasks. I would not use it in production yet: the version is young, and skill self-evolution without testing will sooner or later cause problems in a report to the owner. The rule for managers is simple.

Before choosing a framework, document on one page what data each agent can see, who writes to the shared session, which gateway sends data to the model, and how to roll back a broken skill.

If no one in the company can fill out this page, the framework will not deliver results.

If it can, the framework becomes a replaceable component, and the first measurable result appears within weeks rather than quarters.