Skip to content

Sympozium Roadmap

Last updated: 2026-10-30, after v0.12.1.

This page says where the project is, what is being worked on, and what comes next. It is maintained: when something here ships or is dropped, the page changes in the same pull request. Detailed plans live in the linked design notes and GitHub issues; this is the map, not the territory.

Direction

Sympozium is the coordination layer for agents on Kubernetes: it decides what agents do. Agents and their own model keys, policy and tool gating, ensembles and delegation, channels, schedules, memory and the console, all as Kubernetes resources under your RBAC and GitOps. See Positioning.

Decision of 2026-10-30: isolation and density are an execution plane, and the execution plane is pluggable. Sandboxed, suspendable agent runtimes now exist with far more people behind them than this project has: Google's Agent Substrate (which kagent is rebuilding on) and Kubernetes Agent Sandbox. Sympozium already runs an AgentRun as a Job, as an Agent Sandbox, or as a Celln micro-VM cell. The plan is to add Agent Substrate as an optional backend, keep Celln as the self-hosted, zero-dependency option rather than the default, and spend the effort that frees up on the coordination layer. The reasoning is in Sympozium vs Google AX and Agent Substrate; the work is tracked in epic #692.

Two threads carry on from 0.12 unchanged:

  • Agent-bound model access. An Agent brings its own ModelConnection and key; nothing on an execution node ever holds a model credential, and every model request leaves through the model gateway under that Agent's identity. This holds for every backend, including Substrate. See Agent-bound model access.
  • Bring-your-own harness. External runtimes (Codex, Claude Code, Pi, Hermes, Goose and friends) run under an administrator-approved AgentRuntime with a versioned adapter contract. See the harness epic (#349).

Shipped in 2026

When What Where to read
Oct (0.12) Agent-bound model access: model-free Celln fleet, per-Agent keys pinned to their origin, any OpenAI- or Anthropic-compatible provider, direct Agent API (messages and chat/completions), onboarding wizard design, guide
Oct (0.12.1) In-place gate retry for agent-runner and harness-mode runs (resume), memory update/forget, thinking and maxTokens/temperature tuning, faster Celln admission and turns CHANGELOG
Sep–Oct Celln tenancy: one-shot and enduring runs from any authorised namespace, one-install fleet packaging, trusted transport, operator and tenant runbooks epic #495, fleet installation
Sep Enduring AgentRuns: persistent parents, disposable turn cells, exact-turn cancellation, lease and budget accounting epic #464, design
Aug–Sep Harness mode: AgentRuntime, adapter contract v1alpha1/v1alpha2, reference adapter, conformance suite, run-detail provenance epic #349, harness adapters
Mar–Jul Web console, sympozium doctor, token usage and cost tracking, ensemble delegation (delegate_to_persona, canvas, await/resume), Slack, Discord, Telegram and WhatsApp channels, MCP servers and skill packs Concepts, ensemble roadmap

In progress

  • Pluggable execution plane, Substrate spike (#692). Four weeks with exit criteria: an AgentRun runs as a Substrate actor on the maintainer's cluster, under a SympoziumPolicy, with the Agent's key never inside the sandbox, phases and logs on the AgentRun, and a written go/no-go with latency and idle-cost numbers against the Job and Celln backends. No-go is a valid outcome.
  • Celln becomes an option, not the default (#693). sympozium install deploys the fleet only with --celln-fleet; the Job backend is the default for new Agents; onboarding offers Celln alongside Agent Sandbox and, later, Substrate. Existing fleet installs keep their fleet.
  • Celln turn latency. With the provider out of the picture, a one-shot turn costs about 2.5 s of Sympozium and Celln time (admission 1.5 s, the rest observation and cleanup). Next: cache the resolver's reads during admission, make the per-turn tool budget a tool error instead of ending the turn, and let an Agent lend a leaner toolbox than the whole profile. Tracked in #686; measurements in the design note.
  • Harness epic remainder (#349). Signature or attestation policy for approved runtime images; whether remote MCP credentials stay visible to adapters or are mediated by a local proxy; SkillPack MCP server conformance for Streamable HTTP clients; stable failure categories for one-shot harness runs; pinned examples for Codex, Claude Code and Goose with support tiers; scheduled smoke tests with two real adapters.
  • Ensemble policy and safety (phase 4 of the ensemble roadmap): relationship-scoped delegation rules, cycle detection, timeouts.

Next

Ordered by what unblocks the most users. Each has, or will get, an issue.

  1. Substrate as a released backend (phase 1 of #692, only on a go): spec.backend: substrate, validation and parity tests, install wiring or a documented external install, backend choice in the console, and enduring runs as suspended actors instead of leased parents.
  2. Fleet operations. A per-namespace spend ceiling and refusal events on the run (#688); a KVM CI lane that boots cells on an enforcing CNI and publishes release evidence (#685). On the Celln side: secret providers with live withdrawal, fleet-wide revocation, and a persistent execution model, tracked in its roadmap project.
  3. Observability. OpenTelemetry gen_ai.* semantic conventions on agent-runner spans (maintainer take-over of an external PR, in review), and an OTel transport with durable execution audit retention for the fleet.
  4. Provider breadth without code. Presets for OpenAI-compatible gateways (Databricks, OrcaRouter and others) on top of the custom connection, model discovery per provider where the API allows it, a reachability condition on ModelConnection status (#15), the Responses API for OpenAI reasoning models with tools (#608), and a guide for subscription auth through a Claude Code harness adapter (#684).
  5. Channels. Ensemble-aware routing (a receiver persona and @mention delegation, #257), an agent that can ask a question through its channel and wait for the answer mid-turn (#20), and the long-standing requests for Signal, Teams, Matrix and WebChat, in the order people ask.

Small, well-scoped items carry the good first issue label; the current set is listed on the issues page.

Not planned

  • A model credential held by the fleet, a node, or the installer. 0.12 removed it on purpose; see the threat model in the design note.
  • Agent pods with host networking, host PID, or privileged SkillPacks under an external runtime.
  • Growing Celln's scope. It stays maintained as the self-hosted isolation option (turn latency, fleet operations), but harness-in-a-cell for Pi and Hermes (#687, closed) and other new data paths into the cell are not planned; a sandbox runtime with a community behind it is where that belongs.
  • Competing with Agent Substrate or Agent Sandbox on density or isolation depth. Sympozium sits above them.
  • Reviving the February 2026 OpenClaw gap list as a plan. Its items are either shipped (dashboard, doctor, usage tracking, delegation), folded into the lists above, or not wanted; the original design is kept as history.

How work is tracked

  • Issues are the backlog. Epics carry the epic label and a checklist; slices link back to their epic. Labels name the area (celln, harness, console, channels, provider) and the kind (bug, enhancement, documentation, help wanted, good first issue).
  • Pull requests are reviewed against current main. One untouched for 45 days is labelled stale and closes two weeks later unless someone speaks up; a maintainer may instead take it over, keeping the author's commits and credit. Issues get 120 days and three weeks of grace. pinned exempts an item from both.
  • Releases are cut by release-please from conventional commits; the CHANGELOG is the record of what shipped when. Celln is pinned in config/celln/release.json and moves by its own release.
  • Design notes under Design & History record why; this page records what and when.