Ship
6 current practices.
StandardShipSH-001Stop treating green CI as proof of safetyPassing tests doesn't mean an agent's change is safe. Build systems where deploying is safe by default.
Reviewed Aug 2026ShipSH-002Instrument agents with purpose-built tracing before shippingOne trace per agent run, with child spans for every model call, tool invocation and handoff, and token cost tagged per feature.
Reviewed Aug 2026WatchShipSH-003Ship observability from day one — AI features get one chanceDon't ship an AI feature without observability. When AI tooling breaks, users abandon it for months rather than retrying tomorrow.
One source · a little soon to recommendShipSH-004Version prompts and agents like codeTrack prompt and agent versions so behaviour changes are attributable and reversible.
Reviewed Aug 2026ShipSH-005Stream output, and design the approval gate into the UXStream tokens for perceived responsiveness, and put explicit approval in front of consequential actions.Reviewed Aug 2026WatchShipSH-006Put the agent where the work already happensDeploy agents into existing surfaces like Slack so their work is visible and learnable — and let them push back on bad plans.One source · a little soon to recommend
Reviewed Aug 2026ShipSH-002Instrument agents with purpose-built tracing before shippingOne trace per agent run, with child spans for every model call, tool invocation and handoff, and token cost tagged per feature.Reviewed Aug 2026WatchShipSH-003Ship observability from day one — AI features get one chanceDon't ship an AI feature without observability. When AI tooling breaks, users abandon it for months rather than retrying tomorrow.One source · a little soon to recommendShipSH-004Version prompts and agents like codeTrack prompt and agent versions so behaviour changes are attributable and reversible.Reviewed Aug 2026ShipSH-005Stream output, and design the approval gate into the UXStream tokens for perceived responsiveness, and put explicit approval in front of consequential actions.Reviewed Aug 2026WatchShipSH-006Put the agent where the work already happensDeploy agents into existing surfaces like Slack so their work is visible and learnable — and let them push back on bad plans.One source · a little soon to recommend