The problem

Agents can now use any tool through MCP, but when they need a human they send a paragraph in chat. And every repeat of the same situation costs the same model calls again.

What PRISM does

PRISM is an agent harness that compiles what agents figure out into things that keep running without the model:

• One sentence → a live workspace. The composer samples your MCP tools and emits A2UI v0.9 panels with bindings {server, tool, args, JSONata}. PRISM validates every proposal by executing it, then panels refresh straight from the MCP tools with 0 model calls.
• Watchers → incidents. JSONata conditions run on every refresh. When one fires, PRISM checks for a saved playbook; if none matches, it wakes an event-driven investigate agent that reads the systems and sends a person a decision surface: live evidence panels, one proposed write with exact arguments, and a real dry run. Then the agent’s turn ends — while the decision waits there is no process and no tokens, just a row.
• Human control, enforced in infrastructure. Hold-to-approve (also from a judge’s phone via QR) mints an ES256 token bound to the tool and exact arguments; agentgateway verifies it with CEL before the write reaches the server. Unsigned or edited writes are blocked at the gateway. Kubernetes RBAC is the last line.
• Playbooks: the harness learns. After an approved fix, save it as a playbook. Next time the same incident replays in optimal time (eg. ~25 ms) with 0 model calls (“Same incident, twice: eg. $0.42 → $0.00”).
• It adapts to the person. Send a proposal back with a note (“don’t roll back, set the image to v1”): the same agent resumes from its saved conversation, revises, and the playbook records your fix — “learned from Arnab’s correction”.
• Deep research as evidence. Nimble Web Search Agent runs become live research panels (progress, then a cited report with per-claim confidence and primary/secondary sources).

Sources

Kubernetes (in-cluster MCP server, minikube), an Event DB (hackathon ops: help requests, mentors), GitHub, Stripe (test mode), Sentry (official Sentry MCP behind PRISM’s bridge), Nimble (live web). agentregistry is the discovery layer, agentgateway the governance layer, PRISM the interface layer. PRISM is itself an MCP server: Claude Code can add panels or send decisions.

Observability (Sentry)

Agent runs (invoke_agent → chat claude-opus-5-5 → execute_tool), approvals and playbook replays (no LLM spans) are traced in Sentry; every activity line is a Sentry log; metrics like prism.panel.refresh vs prism.model.calls show “the model rests” in two numbers; agentgateway’s OTLP traces go through an OpenTelemetry Collector to Jaeger and Sentry. Sentry caught a real bug in our harness during the build.

Built today

TypeScript end to end: Next.js + A2UI renderer, Hono API, SQLite, MCP SDK v2 (stateless 2026-07-28), Anthropic SDK (claude-opus-5-5, adaptive thinking, effort medium, prompt caching) routed through agentgateway’s LLM route and agentregistry for storing the agent artifacts.