Astonish

Appearance

Flow Distillation ​

Distillation is the process of converting an exploratory chat session into a deterministic, parameterized flow. The /distill command analyzes the agent's execution trace — every tool call, LLM invocation, and decision point — and produces a reusable YAML workflow.

How It Works ​

When you run /distill, the agent:

  1. Analyzes the execution trace — Identifies the sequence of actions taken during the session.
  2. Detects parameters — Finds values that varied or were user-supplied, converting them into declared parameters.
  3. Identifies decision points — Extracts conditional logic into branching nodes.
  4. Generates the YAML — Produces a complete flow definition with proper node types, edges, and state wiring.
  5. Validates — Runs the flow schema validator to ensure correctness.

Example: Distilling a PR Summary Flow ​

You:    "Get the latest PR from our repo and summarize it for the team"
Agent:  [tool: github_list_prs] Fetching PRs from org/repo...
        [tool: github_get_pr] Got PR #142: "Add rate limiting to API gateway"
        [llm] Here's a standup summary: PR #142 adds rate limiting using a
        token bucket algorithm. Key changes: new middleware in api/gateway.go,
        config options for burst/rate, integration tests.
You:    "/distill"
Agent:  [distill] Analyzing execution trace (3 steps)...

        I've created a flow from this session:

        Name: summarize-latest-pr
        Parameters:
          - repo (string, required) — extracted from "org/repo"
          - channel (string, default: "#engineering")

        Nodes: fetch_prs → get_latest → summarize → post_summary

        Saved as: summarize-latest-pr

        Would you like to edit the flow or schedule it?

The Generated Flow ​

yaml
name: summarize-latest-pr
description: Fetch the latest PR and post a summary to the team
params:
  - name: repo
    type: string
    required: true
  - name: channel
    type: string
    default: "#engineering"

nodes:
  - id: fetch_prs
    type: tool
    tool: github_list_prs
    args:
      repo: "{{repo}}"
      state: open
      limit: 1

  - id: summarize
    type: llm
    prompt: |
      Summarize this PR for a team standup. Be concise.
      PR: {{fetch_prs.output}}

  - id: post_summary
    type: tool
    tool: slack_post
    args:
      channel: "{{channel}}"
      text: "{{summarize.output}}"

edges:
  - from: fetch_prs
    to: summarize
  - from: summarize
    to: post_summary

Using Distillation ​

Distillation is available in multiple contexts:

  • Studio — Type /distill in the chat. The agent previews the identified task and generates the flow immediately.
  • CLI — Type /distill during a chat session. The agent previews the task and asks for confirmation before generating.
  • Agent tool — The distill_flow tool can be called programmatically by the agent (e.g., during scheduled flows).

The generated flow is saved to the platform and appears in your Flows list in Studio.

Team Knowledge in Cloud Deployments ​

In cloud deployments, distilled flows become shared team knowledge:

  • Flows are automatically tagged with the originating session for traceability.
  • Team members can discover flows created by others in the Flows tab.
  • Popular flows surface in search results and recommendations.
  • Forking a team flow creates a personal copy for customization.

This creates a flywheel: the more your team chats with Astonish, the more reusable automation accumulates in your flow library.

Tips for Better Distillation ​

  • Be explicit about inputs — The agent detects parameters more accurately when you clearly state variable values.
  • Keep sessions focused — Distill single-purpose sessions rather than sprawling multi-topic chats.
  • Review the output — Distillation is a starting point. Edit the YAML to add error handling or tighten prompts.
  • Test with different inputs — Run the flow with varied parameters to verify it generalizes correctly.

Next Steps ​