Fixes inside open source agent platforms
Agent platforms break in the seams between the controller, the agent runtime and the model provider. I fixed those seams in four open source projects, each with a test that fails on the main branch.

The problem
On Sympozium, agent pods kept a fixed 1Gi memory limit whatever the runtime declared, so larger agents were killed for memory. The MCP server resource always showed 0 tools, and a rejected schedule retried every 30 seconds with nothing on its status.
On AgentField, legacy requests sent an execution ID to the agent but never wrote the execution row, so status updates got a 404. On UniFi MCP, a relay that registered again kept receiving traffic on its dead WebSocket.
What I built
Reproduced each bug on the main branch and added a regression test with each fix. Two maintainers confirmed the tests fail on main and pass on my branch.
Sympozium, a Kubernetes controller for AI agents in Go. Carried the runtime resources onto the agent pod, filled the MCP tool status through the MCP bridge, stopped the retry loop on admission denials and made the utility images configurable for air-gapped clusters.
AgentField control plane in Go. Wrote the execution row before the legacy handlers reach the agent and copied the final status onto it. After review I kept HTTP 202 replies non-terminal, so background reasoners are not marked done too early.
UniFi MCP on Cloudflare Workers and Durable Objects. Kept the relay bound to the newest socket across rebuilds, with three lifecycle tests.
AnythingLLM collector in Node.js. Fixed the URL check that read YouTube channel links as single videos, with a test file of real links.
How the work ran
Each fix went through public review in its pull request. The AgentField fix changed after the first review, and the merged version includes that rework.
Result
5 pull requests merged in Sympozium in one week.
The AgentField fix merged with 683 lines added and 20 removed across 4 files.
The UniFi MCP maintainer reproduced the bug in the real Workers runtime and ran the full suite of 238 tests green on my branch.
The AnythingLLM fix merged into a project with 67,469 GitHub stars.
“Nice, a schedule spinning every 30s with nothing on its status is exactly the kind of thing that's miserable to debug.”
Evidence
github.com/sympozium-ai/sympozium/pull/664Sympozium PR 664, runtime resourcesgithub.com/sympozium-ai/sympozium/pull/666Sympozium PR 666, MCP tool discoverygithub.com/sympozium-ai/sympozium/pull/665Sympozium PR 665, schedule retry loopgithub.com/Agent-Field/agentfield/pull/1099AgentField PR 1099github.com/sirkirby/unifi-mcp/pull/865UniFi MCP PR 865github.com/Mintplex-Labs/anything-llm/pull/6644AnythingLLM PR 6644
These projects are public, so they are named. Every number links to the pull request or the repository.


