What's new in 0.2.0Copy Anchor Link

0.2.0 is the first release since 0.1.3, and the first measured against other harnesses. It brings the seventeen 0.2.0 pre-releases together for anyone running agents behind their own product: more models, MCP and subagents, a much larger HTTP API, and tighter handling of secrets. A 0.1.3 agent and its store boot on it unchanged.

Measured against Hermes and OpenClawCopy Anchor Link

The same 25 multi-turn tasks, the same tools and the same pinned providers, five runs each, graded on the tools' final state rather than the reply's wording:

• Ahead of OpenClaw 2026.10.5 on all 11 models, and level with or ahead of Hermes 0.19.0 on 8 of 11.
• 125/125 on five models: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, GLM 5.3 and DeepSeek V4 Pro.
• Cheapest of the three on 10 of 11 models. Hermes cost $0.01 less on GPT-6.1 Sol.
• Behind Hermes on three: Gemini 4 Argon, qwen3.5-9b and gpt-oss-20b. These are listed with the rest.

The full tables, the dialect comparison and how the runs were made are on Benchmarks. Every run is committed under evals/harness/.

HighlightsCopy Anchor Link

Agents that finish the job. init now writes native tool calls wherever the model supports them; on the benchmark native tied or beat the plain-text dialect on all 13 models measured both ways. A step that would end silent, on "I'll do that", or claiming a save that never happened gets one note from the runtime and one more step. A retry the tool itself asked for is no longer refused as a loop, and a tool result can take a quarter of the context window instead of a flat 2,000 tokens.

Any model, through more than one door. Model transports are pluggable: AWS Bedrock ships (api: bedrock-converse), and a plugin can register another. fallbacks tries the next model when an endpoint fails before any output. Images reach models with vision, signed thinking is replayed across tool calls, and a refused reply ends the turn as model_refused instead of an empty answer.

Prompt caching that survives a memory write. Only byte-stable context is sent as system messages, so a saved note no longer empties OpenAI's cache, and Claude through OpenRouter gets cache markers.

MCP and subagents. tools.providers.mcp serves tools from remote MCP servers over Streamable HTTP, pinned by name like any other tool. subagents: routes a call to a throwaway child of the same agent, optionally on a cheaper model, and the parent reads its artifact instead of the raw output.

An HTTP API for running other people's agents. Agent templates, per-agent secrets, metered model calls (/v1/usage) with ids a biller can debit, signed outbound webhooks, concurrency and token limits, exact-text deliveries, repair of an agent that will not load, and GET /v1/backup for the whole silo. dispach/client ships its own type declarations and no terminal UI.

Secrets stay secret. file_read and grep never read .env, key files or credential directories, path rules match the normalised path, exec can run with the runtime's secrets scrubbed, and X_FILE sets a variable from a file without it entering the environment.

Memory without a restart. A saved note, or an edit to MEMORY.md, reaches every session on the next turn. memory_forget deletes exactly the notes named, and a conversation can be taken out of recall without deleting it.

Runs where you deploy. A Kubernetes guide and example, a drain on stop for running turns, ffmpeg in the image for voice notes, and one host token shared by every agent in a sandbox.

Upgrading from 0.1.3Copy Anchor Link

1. Back up store.db. Migrations 18 to 31 run on first boot and have no down-migration.
2. Install 0.2.0 (npm install --global dispach@latest, brew upgrade dispach, or docker pull ghcr.io/moeen-mahmud/dispach:0.2.0) and start it. No agent.yaml change is needed.
3. Check the four breaking changes, if they touch you:
   • a plugin must declare dispachApi: "^0.2";
   • a reload during a turn answers 202 pending instead of 409;
   • GET /v1/agents/:id on a stopped agent returns a status: "disabled" row instead of 404;
   • a scheduled run that ends in timeout, max_steps, no_progress or error is recorded as an error.
4. Expect one prompt-cache miss on the first turn, because workspace files now reach the model under their names.

An existing manifest keeps its tool-call dialect. To move an agent to native tool calls, set tools.dialect: native.

ExperimentalCopy Anchor Link

These are on the HTTP API only and may change within 0.2.x: rooms, participants and DMs; stand-ins and delegation between members' agents; team memory scopes; per-message model roles; the Slack, Microsoft Teams and A2A channels; media (transcription, image generation, spoken replies); and Firecrawl as a web backend. See Stable and pre-releases for what experimental means here.

Read moreCopy Anchor Link

• CHANGELOG.md: every change, one line each.
• docs/00-DECISIONS.md: why each of them was made, with its measurement.
• evals/CLAIMS.md: every public number and the file it comes from.