Test an AI Agent
Run an isolated test in Preview to verify reasoning and tool calls before sending live writes. Both Preview and Live test use the real model and bill Glow AI tokens.
- Use it to
- Test in Preview (read-only tools run, writes Proposed) or Live test, approve tool calls, and configure memory
- Find it
- AI Agent → Test & Debug → Preview or Live test → Run Test
- Cost
- Preview and Live test use Glow AI tokens (plus Step credits for executed actions)
- Limits
- Preview does not read or save memory · Shape repair allows 1 extra AI call
- Output
- {{ N.result }} (or {{ N.result.<field> }}) and {{ N.estimatedCreditsUsed }}
Preview or Live test
- Runs the real model and read-only actions (find email, list users).
- Marks write actions and external MCP tools as Proposed · not sent.
- Starts fresh without reading or saving Conversational Memory.
- Uses model credits with no confirmation dialog.
- Dispatches real app actions and external MCP tools.
- Pauses on approval-gated actions until you Approve or Reject.
- Reads and saves Conversational Memory for the active key.
- Prompts once (Switch to Live test?) and uses model and Step credits.
Switch back to Preview. The toggle stays on Live test in your open tab until you switch it back.
What the agent works on
Tests resolve Instructions using the latest output or Testing data of earlier Steps. Edit an upstream Step’s Testing data (Test & Debug → Testing data) or paste sample text in the prompt.
Approve or reject an action
When a Live test or run hits an approval-gated tool:
Open the pending decision
Open Executions on the agent to see 1 action needs a decision before the agent can continue.
Inspect prepared inputs
Expand Prepared inputs to check recipients and fields. Secret values are hidden.
Approve or reject
Click Approve, Reject, or Open in Approvals to continue.
Read the result
Inspect the Trace
Each test logs three sections in Test & Debug:
- Output: Final answer as Formatted or Raw JSON (
{{ N.result }}or{{ N.result.field }}). - Agent activity: Model and tool calls with status (
Succeeded,Proposed,Awaiting approval,Failed) and timing. - Detailed usage & metrics: Tokens and
{{ N.estimatedCreditsUsed }}.
| Activity label | Meaning |
|---|---|
| Search the web / Read a web page / Run code | Built-in helper call. |
| Recall this conversation / Search reference files | Context lookup in memory or attached files. |
Run workflow: {name} / {MCP connection} · {tool} | Subflow call or external MCP tool call. |
| Look up values: … | Option lookup, not a write (details). |
Choose which runs share memory
Conversational Memory is off by default. Enable it in Context or on the Memory tile (Glow-managed Gemini models) and set Memory Instructions.
Memory is scoped to the workspace, automation, and Agent Step:
| Memory key option | Best for | Who shares context |
|---|---|---|
Earlier-Step field ({{ 1.threadId }}) | Support tickets, chat threads, per-customer | Runs with the same thread, ticket, or customer ID |
| Share one conversation across all runs | Personal assistants or team digests | Every run of this Agent Step (default key) |
| Custom memory key | Composite keys ({{ 1.org }}-qa) | Runs whose resolved custom key matches |
Missing required memory IDs stop the Step before calling AI.
Isolate customer context. Never share one static Conversation ID across customers. Leave Conversational Memory off when unused.
Upgrade earlier shared memory
Older Agent Steps show Upgrade Memory, which starts clean memory without importing prior history.
Credentials and conversation memory
Glow blocks workspace Secrets and API keys ({{ $secret.* }}, {{ $key.* }}) from AI prompts and strips known secrets from memory. Authenticate via connected accounts or HTTP Request auth fields (variable safety).
Check the answer shape
In Output, Separate fields validates the reply against your schema (What it passes on). Ask the AI to fix a wrong shape allows one extra model call to repair invalid JSON without re-running tools ({{ N.result.field }}).
Recover from a failed run
Check Executions to see which tool calls finished, and wire an Error Path to a fallback (Error Handling).
Check external writes before retrying
Resuming an approved call verifies dispatch state and ownership so that action never runs twice. A new run does not undo writes and can duplicate external actions, so check the external app first.
Error branches do not check reasoning. Require approval in Tools for agent writes, or add a Human Review Step before downstream writes.
Example configurations
Adapt these prompts and Output → Separate fields (Advanced (JSON Schema)) schemas. Boolean fields like requires_human_review are plain data and do not pause a run by itself.
B2B Prospect Research
B2B Company & Prospect Research
Uses Search the web and Read a URL you give it on public URLs (Identity: B2B researcher; Dos/Don'ts: check company site first, return null when unknown).
Research "{{ 1.email | extract_domain }}" with web search.
Find value proposition, target customer, headcount, and 90-day news.
Email: {{ 1.email }} · Lead: {{ 1.name | capitalize }}{
"type": "object",
"properties": {
"company_name": { "type": "string" },
"industry": { "type": "string" },
"value_proposition": { "type": "string" },
"target_market": {
"type": "string",
"enum": ["SMB", "Mid-Market", "Enterprise"]
},
"recent_news": { "type": ["string", "null"] },
"recommended_outreach_angle": { "type": "string" }
},
"required": [
"company_name",
"industry",
"value_proposition",
"target_market",
"recommended_outreach_angle"
]
}- Run
- Open AI Agent → Tests tab, keep Preview selected, and click Run test
- You should see
- Read-only tools run live, writes stay Proposed, and the trace logs tool calls and tokens.
What’s Next?
- Debug a Run: trace a failed Step and inspect its results.
- Human Review: require approval before a downstream action.
- AI Agent: change the agent’s tools, instructions, or Answer format.