# Limits

One table of the numbers you can hit, each linked to the page that explains it.

These are the platform's fixed limits and defaults. Each row links to the page
that explains what to do about it. Contact support@introspection.dev to raise
a plan limit.

## Organization and identity

| Limit | Value | Where |
| --- | --- | --- |
| Members per organization (team members, end users, anonymous IDs, and each Slack channel or DM) | Free and Pro 100, Max 1,000, Enterprise 10,000. At the cap, creating a member returns `409`; a new Slack channel's messages are dropped. | [Serve many customers](/patterns/multi-tenant) |
| `user_id`, `anonymous_id`, `conversation_id` length | 255 printable characters, no whitespace | [Add to your app](/patterns/add-to-your-app) |
| Service-account access token | 60 minutes; no refresh token | [Credentials](/identity/credentials) |
| Runner session (`ttl_seconds`) | 1 hour by default | [JavaScript](/reference/javascript) |
| Native app access token / refresh token | 60 minutes / 30 days, rotating | [Authentication](/reference/authentication) |
| User connection authorize link | 600 s by default, 60 to 86,400 s; Pipedream links at most 4 hours | [Per-user tool access](/patterns/per-user-tools) |

## Tasks and runs

| Limit | Value | Where |
| --- | --- | --- |
| Idle timeout between turns | 600 s by default, and 600 s is also the maximum; `0` tears down after provisioning. Not counted while streaming or `awaiting_user`. | [Task operations](/reference/task-operations) |
| Turn duration | No per-turn cap | [Task operations](/reference/task-operations) |
| Inactive task | After 10 minutes idle the sandbox is torn down; after 3 hours with no activity the task stops holding a concurrency slot. The conversation stays, and the task can be prompted again. | [Ask the user](/patterns/ask-the-user) |
| Interrupt `expiresAt` | Recorded only; nothing enforces it | [Ask the user](/patterns/ask-the-user) |
| Subagents | 64 active runs per task across all nesting levels; a further `start` fails until one finishes (no queue). Nesting depth 2. | [Multi-agent](/patterns/multi-agent) |
| Concurrent tasks per organization | Set by your plan, capped by the data plane's size; extra tasks wait as `queued`. Contact support to raise it. | [Task operations](/reference/task-operations) |
| Concurrent eval trials per organization | Set by your plan, in a separate lane from tasks (Harbor and `kind: eval` trials; local `eval run` uses neither) | [Harbor](/improve/harbor) |
| Time a task waits in the queue | About 1 hour (checked hourly), then it is cancelled | [Trigger your agent](/patterns/trigger) |
| API requests per organization | No rate limit by default. `429` means a run's stream isn't ready yet (wait for `Retry-After`; the SDKs do), too many sign-in codes, or a per-plane limit your deployment turned on. | [Add a chat to your web app](/patterns/web-chat) |
| Spend | No spend or budget caps; watch cost with an [alert](/observe/metrics#alert-on-a-threshold) | [Metrics](/observe/metrics) |
| Agent-scheduled follow-ups | 100 per member; more returns `409` | [Let your agent schedule its own follow-ups](/patterns/agent-follow-ups) |
| Experiment arms | 2 to 20 per experiment | [Experiments](/improve/experiments) |

## Files and recipes

| Limit | Value | Where |
| --- | --- | --- |
| Files attached to a task | 100 per task | [Send files in, get files out](/patterns/files) |
| File attached to a task, or saved from it as an output or memory | 25 MB each | [Send files in, get files out](/patterns/files) |
| File uploaded through the Files API or CLI | 50 MB each | [Send files in, get files out](/patterns/files) |
| File shipped in the recipe's `files/` | 25 MB each | [Manifest](/build/package) |
| Recipe patch uploaded by `introspection dev` | 10 MiB, the maximum (`--runtime-max-total-bytes` can only lower it) | [Local development](/build/local) |
| Deployed recipe total size | No cap | [Manifest](/build/package) |
| `runtime.resources` | Can only raise the deployment floor; at most 100 CPU and 200Gi | [Deploy your agent](/ship) |
| Slack file fetched by the agent (`fetch_file`) | 100 MiB to the sandbox disk; the 25 MB cap still applies to outputs | [Put your agent in Slack](/patterns/slack) |

## Reads, retention, and export

| Limit | Value | Where |
| --- | --- | --- |
| Conversations and telemetry | Kept unless retention is enabled for your data plane; set per plane, off by default | [Your data](/security/data) |
| Audit log | At least 2 years | [Audit log](/security/audit) |
| Tasks, files, memory, members | Until deleted | [Memory and deleting data](/patterns/memory-and-data) |
| Metrics query window | 400 days | [Metrics](/observe/metrics) |
| `conversations get --ids-file` | 20 IDs per file; IDs named on the command line have no cap | [Conversations](/observe/conversations) |
| `conversations export --lookback-days` | 1 to 365 days of records within one conversation's export | [Evals CLI](/reference/cli/evals) |
| `tasks list --limit` | 1 to 1,000 (default 100) | [CLI deploy commands](/reference/cli/deploy) |

## SDK clients

| Limit | Value | Where |
| --- | --- | --- |
| HTTP retries | 2 (JS and Python). `429` on any method, `502`/`503`/`504` on GET only, never multipart; backoff from 500 ms. | [JavaScript](/reference/javascript) |
| Stream recovery window | 300 s (JS `timeoutMs: 300000`, Python `timeout=300.0`), 5 reconnects. Renewed by new content; when it runs out only the client stream fails. | [JavaScript](/reference/javascript) |
| Python per-request timeout | 30 s (JS has none) | [Python](/reference/python) |

## Next

- [Task operations](/reference/task-operations)
- [Serve many customers](/patterns/multi-tenant)
- [Memory and deleting data](/patterns/memory-and-data)

Limits

One table of the numbers you can hit, each linked to the page that explains it.

These are the platform’s fixed limits and defaults. Each row links to the page that explains what to do about it. Contact support@introspection.dev to raise a plan limit.

Organization and identity

Limit	Value	Where
Members per organization (team members, end users, anonymous IDs, and each Slack channel or DM)	Free and Pro 100, Max 1,000, Enterprise 10,000. At the cap, creating a member returns `409`; a new Slack channel’s messages are dropped.	Serve many customers
`user_id`, `anonymous_id`, `conversation_id` length	255 printable characters, no whitespace	Add to your app
Service-account access token	60 minutes; no refresh token	Credentials
Runner session (`ttl_seconds`)	1 hour by default	JavaScript
Native app access token / refresh token	60 minutes / 30 days, rotating	Authentication
User connection authorize link	600 s by default, 60 to 86,400 s; Pipedream links at most 4 hours	Per-user tool access

Tasks and runs

Limit	Value	Where
Idle timeout between turns	600 s by default, and 600 s is also the maximum; `0` tears down after provisioning. Not counted while streaming or `awaiting_user`.	Task operations
Turn duration	No per-turn cap	Task operations
Inactive task	After 10 minutes idle the sandbox is torn down; after 3 hours with no activity the task stops holding a concurrency slot. The conversation stays, and the task can be prompted again.	Ask the user
Interrupt `expiresAt`	Recorded only; nothing enforces it	Ask the user
Subagents	64 active runs per task across all nesting levels; a further `start` fails until one finishes (no queue). Nesting depth 2.	Multi-agent
Concurrent tasks per organization	Set by your plan, capped by the data plane’s size; extra tasks wait as `queued`. Contact support to raise it.	Task operations
Concurrent eval trials per organization	Set by your plan, in a separate lane from tasks (Harbor and `kind: eval` trials; local `eval run` uses neither)	Harbor
Time a task waits in the queue	About 1 hour (checked hourly), then it is cancelled	Trigger your agent
API requests per organization	No rate limit by default. `429` means a run’s stream isn’t ready yet (wait for `Retry-After`; the SDKs do), too many sign-in codes, or a per-plane limit your deployment turned on.	Add a chat to your web app
Spend	No spend or budget caps; watch cost with an alert	Metrics
Agent-scheduled follow-ups	100 per member; more returns `409`	Let your agent schedule its own follow-ups
Experiment arms	2 to 20 per experiment	Experiments

Files and recipes

Limit	Value	Where
Files attached to a task	100 per task	Send files in, get files out
File attached to a task, or saved from it as an output or memory	25 MB each	Send files in, get files out
File uploaded through the Files API or CLI	50 MB each	Send files in, get files out
File shipped in the recipe’s `files/`	25 MB each	Manifest
Recipe patch uploaded by `introspection dev`	10 MiB, the maximum (`--runtime-max-total-bytes` can only lower it)	Local development
Deployed recipe total size	No cap	Manifest
`runtime.resources`	Can only raise the deployment floor; at most 100 CPU and 200Gi	Deploy your agent
Slack file fetched by the agent (`fetch_file`)	100 MiB to the sandbox disk; the 25 MB cap still applies to outputs	Put your agent in Slack

Reads, retention, and export

Limit	Value	Where
Conversations and telemetry	Kept unless retention is enabled for your data plane; set per plane, off by default	Your data
Audit log	At least 2 years	Audit log
Tasks, files, memory, members	Until deleted	Memory and deleting data
Metrics query window	400 days	Metrics
`conversations get --ids-file`	20 IDs per file; IDs named on the command line have no cap	Conversations
`conversations export --lookback-days`	1 to 365 days of records within one conversation’s export	Evals CLI
`tasks list --limit`	1 to 1,000 (default 100)	CLI deploy commands

SDK clients

Limit	Value	Where
HTTP retries	2 (JS and Python). `429` on any method, `502`/`503`/`504` on GET only, never multipart; backoff from 500 ms.	JavaScript
Stream recovery window	300 s (JS `timeoutMs: 300000`, Python `timeout=300.0`), 5 reconnects. Renewed by new content; when it runs out only the client stream fails.	JavaScript
Python per-request timeout	30 s (JS has none)	Python