Agent action log

What is the action log?

Every time a Autopilot agent runs, the runtime records what happened. The result is a structured, queryable log that lives in your app as the AgentAction collection. If something is in doubt, the action log is the source of truth. The log captures four things for every run:

The trigger. Which playbook fired and why.
The observation. What the agent looked at (records, signals, prior actions).
The reasoning. What the agent decided and why.
The outcome. Which tools ran, what they returned, what happened next.

This isn’t optional. The log is always on. The runtime cannot run a tool without writing to it.

Why every action is logged

Trust. Users approve drafts faster when they can see the full reasoning chain. Skeptics become believers when they can audit every step.
Debugging. When a playbook misbehaves, the log shows where the reasoning went sideways.
Compliance. Regulated environments need provable audit trails. The log is structured for export.
Graduation. Risk tier graduation uses approval rates from the log. No log, no graduation.
Improvement. Approval and skip patterns reveal which objectives are tuned correctly and which need editing.

Schema

The AgentAction collection is a system-managed collection in every Gainable app. You can query it through the Copilot, the dashboard, or the action log UI in Autopilot.

Field	Type	What it stores
`id`	string	Unique action ID
`agent_id`	string	The agent that produced the action
`playbook_id`	string	The playbook that ran
`objective_id`	string	The objective the playbook was working toward
`scope`	enum	`app_wide` or `personal`
`trigger_type`	enum	`schedule`, `data_change`, `webhook`, `user_triggered`
`trigger_payload`	json	Full payload of the trigger
`observation`	json	Records and signals the agent considered
`reasoning`	text	Human-readable reasoning chain
`tool_calls`	array	Each tool the playbook called, with inputs and outputs
`outcome`	enum	`approved`, `skipped`, `auto_executed`, `rate_limited`, `error`, `pending`
`owner_user_id`	string	The user the action’s draft was routed to
`created_at`	timestamp	When the run began
`resolved_at`	timestamp	When the action reached a terminal outcome
`parent_action_id`	string	If this action was a follow-up to another

Outcomes

Outcome	What it means
`pending`	Draft is waiting in Autopilot
`approved`	Owner approved; underlying tool ran
`skipped`	Owner skipped the draft
`auto_executed`	Action ran without approval (auto tier or auto-with-undo)
`rate_limited`	Action was blocked by a rate limit
`error`	Tool call failed; reason captured in `tool_calls[*].error`

Querying the log

The action log is just a collection, so you can query it the same way you query anything else.

From Copilot

Ask the Copilot directly:

Show all SLA-enforcer actions from the last 7 days
where the outcome was skipped.

Which playbook produced the most rate_limited
outcomes this month?

From a dashboard

Drop a chart or table view onto a dashboard scoped to the AgentAction collection:

Add a chart to the operations dashboard showing
SLA-enforcer outcomes (approved, skipped, error)
over the last 30 days, grouped by week.

From a playbook

Yes, agents can read their own action log. This is how follow-up playbooks know what’s already been tried:

- tool: query_collection
  collection: AgentAction
  filter: "playbook_id = 'stalled_deal_chaser' AND owner_user_id = {{ deal.owner }} AND created_at > now - 7_days"
  limit: 1

If a stalled-deal playbook sees that it already drafted a nudge for this deal in the last 7 days, it can skip drafting another. Ackermann-class loops are bounded.

What gets logged for each tool

Tool	Logged inputs	Logged outputs
`query_collection`	Filter, sort, limit, fields	Record IDs returned, count
`update_records`	Filter, set clause	IDs updated, prior values for undo
`send_email`	Recipient, subject, body hash	Provider message ID, delivery status
`send_slack`	Channel, message hash	Slack message ID
`send_sms`	Recipient hash, body hash	Provider message ID
`notify_user`	User ID, title, body	Notification ID
`call_external`	URL, method, body hash	Status code, response hash
`ask_human`	User ID, question, options	Answer, response time
`create_task`	Task fields	Created task ID
`draft_for_approval`	Underlying payload	Draft ID, owner

Bodies of outbound messages are hashed by default for compactness. The full content is retained on the draft itself in Autopilot until the configured retention window closes.

Retention

Data	Retention
Action records	Indefinite (or per-org configured)
Full draft bodies	90 days after `resolved_at` (configurable)
Tool input/output payloads	30 days (configurable)
Hashes and metadata	Same as the action record

For regulated environments, retention can be raised to 7+ years. For consumer-facing apps with privacy obligations, retention can be tightened.

Common queries

Approval rate per playbook

Show approval rate per playbook over the last 30 days,
sorted highest to lowest.

This is the number you watch when deciding whether to graduate a playbook to a higher risk tier.

Why was this draft created?

Open the draft in Autopilot. The reasoning and observation fields render alongside the proposed action. No need to query the log directly.

What did this agent do this morning?

Show all actions from the daily-briefing agent
in the last 24 hours.

Where are skips concentrated?

Which objectives have the lowest approval rate this month?

A low approval rate means the objective is producing drafts that don’t match what users actually want. Tighten the trigger filter or revise the template.

Best practices

Add a log dashboard early

A dashboard with approval rate, skip reasons, and rate-limited counts is the single most useful artifact for tuning agents. Add it during the autonomy build phase.

Watch errors aggressively

Errors in tool_calls are usually a sign of upstream changes (renamed fields, missing API keys, deleted records). Spikes mean something broke.

Use parent_action_id for follow-ups

When playbook B runs because playbook A’s draft was approved, link them. The thread of reasoning becomes traceable.

Don't write to AgentAction yourself

The runtime owns the schema. Custom apps that try to insert their own action records will fail validation. Use create_task for human-visible TODOs instead.

Learn more

Tools

What gets logged per tool

Risk tiers

How the log feeds graduation

Inbox

Where the log surfaces in the UI

Scopes

How log slicing maps to scope

Get started

Concepts

Prompt guide

Building apps

Gaia Copilot

Gaia Autopilot

Skills reference

MCP connector

Reference

What is the action log?

Why every action is logged

Schema

Outcomes

Querying the log

From Copilot

From a dashboard

From a playbook

What gets logged for each tool

Retention

Common queries

Approval rate per playbook

Why was this draft created?

What did this agent do this morning?

Where are skips concentrated?

Best practices

Learn more

Tools

Risk tiers

Inbox

Scopes

​What is the action log?

​Why every action is logged

​Schema

​Outcomes

​Querying the log

​From Copilot

​From a dashboard

​From a playbook

​What gets logged for each tool

​Retention

​Common queries

​Approval rate per playbook

​Why was this draft created?

​What did this agent do this morning?

​Where are skips concentrated?

​Best practices

​Learn more

Tools

Risk tiers

Inbox

Scopes

What is the action log?

Why every action is logged

Schema

Outcomes

Querying the log

From Copilot

From a dashboard

From a playbook

What gets logged for each tool

Retention

Common queries

Approval rate per playbook

Why was this draft created?

What did this agent do this morning?

Where are skips concentrated?

Best practices

Learn more