August Product Update

Here's everything Plural shipped in August.

🔒Agent Governance: Policies for runs, tools, and spend

• Tool and stack policies. Define policies once at the project level and attach them to workbenches or stacks — each one is evaluated against the exact input of an agent's tool call, so governance rules run before the agent acts instead of after.
• AgentRuntimePolicy. A new CRD centralizes who can start agent runs on a given runtime, restricting execution to designated users or groups.
• Workbench-level budgets. Building on last month's per-mode caps, each workbench now carries its own token or dollar bucket with a minimum-free floor and a live remaining balance.

🔧Workbenches Enhancements: Pi agent support, prompt queues, and custom MCP servers

• Pi agent harness. Pi joins Claude Code, Codex, Gemini, and OpenCode as a supported coding-agent harness for agent runs, with AWS SigV4 auth built into the AI proxy so runs can hit Bedrock directly from the pod.
• Post-PR follow-ups. Workbench jobs now support robust follow-up prompts after the agent opens a PR — queue a follow-up from your CI pipeline, for example, to confirm the deployment is healthy once the change rolls out.
• Cost recommendations to PRs. Scaling recommendations from cost management can now be handed to a Workbench, which turns the recommendation into a reviewable PR.
• Custom MCP servers on runtimes. Attach any remote MCP server to an agent runtime and its tools become available to coding agents, alongside a default model and provider override that persists per runtime.

⚡Platform & GitOps: Operator performance and config flexibility

• Durable operator caches. The deployment operator's caches now persist across restarts, with a configurable storage directory and flush interval — no more re-fetching and re-computing everything after a pod bounce.
• Deferred startup polling. Operators that have been running for over an hour can skip the immediate poll on restart, cutting unnecessary load at startup across large fleets.
• Python file schemas. Helm configuration can now be generated from Python scripts, so you can compute overlays from Git state or other metadata instead of hand-maintaining values files.
• Multiple ServiceContext sources. A ServiceContext can now merge an ordered list of ConfigMap sources, with later sources overwriting earlier values.
• SCM token rotation. SCM connections automatically reconcile when their token Secrets rotate, so credential rotations no longer break git integrations.

🏢Company Updates

We're attending KCD in SF on September 1st. Let us know if you're planning to attend!