v0.1.56 — Build a production agent step by step, and own the harness

Released 2026-06-20. Multi-platform signed/notarized DMG (Mac Apple Silicon + Intel) + EXE + AppImage. Auto-update from v0.1.55 happens on next launch.

The Agent Foundry was powerful but it didn’t tell you where to start. v0.1.56 turns its empty state into a guided, step-by-step builder: at every step it says what you’re building, which step is next, and lets you jump back to any earlier step to edit. Along the way it quietly decodes the buzzwords you hear every day — harness engineering, context engineering, skills, knowledge, MCP, sub-agents — on the exact step where each one lives.

The through-line is ownership: CodeNow helps you build and own your own harness — a unique multi-model, multi-vendor agent — instead of getting locked into a one-click “managed agent” where it’s easy but you don’t see inside and don’t really own anything.

A 9-step guided build

Open a project with no agent yet → the Agents panel now walks you through nine steps, with a clickable progress rail so you always know where you are and can go back to edit:

1. Choose your harness — pick the SDK that becomes your agent’s spine (Claude Agent SDK · Google ADK 2.0 · Cursor SDK · Microsoft Agent Framework), each with a plain-language “best for” and a lock-in badge.
2. Orchestrator — CodeNow scaffolds the real agent file (agent.ts / agent.py) from your chosen kit. Runnable today, yours forever. (Prefer plain language? Describe the goal and CodeNow drafts and writes it.)
3. Model & Reasoning — multi-model with your own keys, plus a fallback so one provider’s outage can’t take your agent down.
4. Knowledge & Memory — ground the agent in your own documents, data, and indexed codebase (RAG) so it answers with facts a generic model can’t know.
5. Tools & Connections — give it real actions over MCP, the open standard that keeps your connectors portable instead of locked to one platform’s catalog.
6. Sub-agents — add helpers the orchestrator can delegate focused tasks to, and keep adding more as the job grows.
7. Guardrails & Oversight — visual limits on what the agent can touch and do, sent as a system prompt on every run, with per-run TrustScores from TrustModel.ai (auto-provisioned, no key to paste) and a human kept in the loop.
8. Observe & Test — every run captured to a replayable, forkable record, and evals that score the agent across models so you can prove what works.
9. Deploy & use daily — use it right here in CodeNow, publish the harness to Agent21, or ship it to a cloud you control.

The buzzwords, decoded where they live

Every step opens with one plain sentence a non-developer gets instantly — what is a harness, what is context engineering, what is MCP, what is RAG — tied to the block you’re building. No glossary to go read; the explanation is on the step.

Own your harness, don’t rent one

Step 1 makes CodeNow’s stance explicit. A one-click “managed agent” (Google’s Antigravity Managed Agents are the clearest 2026 example — “manage the mission, not the machine”) is fast: you POST a config and the vendor provisions and runs the agent in their sandbox. The honest trade-off is that you don’t see inside the harness, you don’t own the orchestration IP, and the managed tier runs on the vendor’s own model — so switching later means rebuilding.

CodeNow takes the other path: an ownable, multi-model, multi-vendor harness whose code, tools, and logic are yours. That ownership is also what makes real human oversight possible — when the harness is yours and not a black box, you can see every run, train it, set guardrails, and override it. Inline guidance on the step helps you choose between Google ADK 2.0 / Antigravity and the Anthropic Claude Agent SDK honestly, without pushing you toward a tier that locks you in.

Deploy where you have credits

The final step covers daily use and deployment honestly — real where we automate it, real commands where you run them:

Each cloud shows the real, copy-to-terminal commands and its actual free-credit program. You own the image and the runtime — there’s no managed-agent box in the loop.

Honest about what’s real today

No mocks. Steps 1–2 create real code (scaffold a runnable orchestrator, or describe-and-draft one). Steps 3–9 explain each block and route you straight into the already-shipped surfaces — the multi-model studio, Tools, Knowledge, Business-mode Guardrails, Runs/Replay, Evals, and Publish — rather than faking an inline editor that doesn’t exist yet. Capturing each block’s config inline, and native drag-to-compose sub-agents, are the next phase.

Upgrade notes

Auto-update from v0.1.55 happens on next launch. No migration steps.

Under the hood

Renderer-only change, gated the same way every release is: full smoke (240+ cases) plus both end-to-end journeys must pass on the tagged commit before anything is built, signed, or published. The first build of this tag was blocked by a flaky browser-navigation step in the journeys, re-run clean, then shipped — a version can’t go out red.

What’s next

• Inline per-block config — let steps 3–9 capture each block’s settings in the wizard, not just route out to the shipped surface.
• Native sub-agent composition — drag-to-compose orchestration via Google ADK 2.0 Collaborative Workflows.
• Windows code signing via Azure Trusted Signing — wired and waiting on account setup.