Limits and Defaults

Agent resource ID

Up to 63 characters

Derived from the agent’s name. A lowercase slug that starts with a letter, ends with a letter or digit, and contains letters, digits, and hyphens. Can’t change after creation. See Create an Agent.

Agent display name

Up to 128 characters

See Create an Agent.

Agent description

Up to 1,024 characters

See Set Up a Self-Managed Agent.

MCP server name

1 to 63 characters

Lowercase letters, numbers, and hyphens only, starting with a letter. Used as the URL path segment. Can’t change after creation. See Create an MCP Server.

MCP server description

Up to 256 characters

See Create an MCP Server.

Guardrail name

1 to 63 characters

Lowercase letters, numbers, and hyphens only. Used in the resource path. Can’t change after creation. See Create a Guardrail.

Guardrail display name and description

Display name up to 253 characters, description up to 1,024 characters

See Create a Guardrail.

LLM provider display name

Up to 253 characters

Agentic Data Plane derives the resource ID from it. See Configure an LLM Provider.

Tags on an agent, MCP server, LLM provider, or guardrail

At most 50 key/value pairs, each value up to 256 characters

See Track Spend by Tag and Create an MCP Server.

System prompt

Up to 50,000 bytes of UTF-8

Applies to the agent and to each subagent’s instructions. That is 50,000 characters of unaccented Latin text and fewer with accented, non-Latin, or emoji characters. See Create an Agent.

Subagents per agent

At most 16

Each subagent needs a description of 1 to 1,024 characters, which the parent agent reads to decide when to delegate. See Create an Agent.

Model identifier

1 to 128 characters

Applies to the agent’s model and to a subagent’s model override. See Create an Agent.

MCP servers per agent

At most 32 references

Counts the root agent’s servers plus every subagent’s servers, as references rather than distinct servers. See Create an Agent.

Tool call timeout

10 minutes

A managed agent waits up to 10 minutes for each tool call, including a call to a subagent. The limit is fixed. See Troubleshoot Agents.

Max iterations

Default 100 when unset or 0, at most 200

The upper bound on reasoning loops per request. The API rejects a value above 200. No UI control sets it. Set it with rpk ai agent update --max-iterations, the API, or an applied configuration file. See Max iterations.

Self-managed agent client secret lifetime

Default 90 days, no maximum enforced

Applies when you create a secret without an expiry, which is what the Credentials tab always does. To set a different lifetime, create the secret with rpk ai agent credential create --ttl. Every secret has an expiry. There is no never-expires option. The plaintext is shown only at creation. See Set Up a Self-Managed Agent.

Client secrets per self-managed agent

10 unexpired secrets

Only secrets that have not expired count toward the cap, so an expired secret frees its slot on its own. A request that would exceed the cap fails with resource_exhausted and the message secret cap exceeded. Revoke a secret you no longer use to free its slot immediately. See Set Up a Self-Managed Agent.

Span export token

About an hour

Mint tokens as you export rather than pinning one in an environment variable. See Self-Managed Agent Telemetry Reference.

Schedule triggers per agent

Up to 10

Each with its own input prompt. See Run an Agent on a Schedule.

Gap between scheduled runs

At least 15 minutes

Agentic Data Plane rejects a schedule whose runs fall closer together as running too frequently. See Run an Agent on a Schedule.

Runs in flight per trigger

One

A run that stays in Running for more than about 15 minutes moves to Unknown. See Run an Agent on a Schedule.

Microsoft Teams trigger health check

About every 30 seconds

The messaging bridge validates the trigger’s credentials with Microsoft in the background. See Connect an Agent to Microsoft Teams.

Code mode script

64 KiB

The code passed to execute. See Code Mode.

Tool calls per execute

At most 50

One code mode execution. Each sandbox also has a memory limit and a runtime limit. See Code Mode.

List filter expression

Up to 1,024 characters

The AIP-160 filter on the list APIs for MCP servers and OAuth providers. order_by takes up to 256 characters. See Create an MCP Server.

JavaScript server tools

At most 64 per bundle

Tool names match ^[a-zA-Z0-9_-]{1,64}$ and must be unique. See JavaScript.

JavaScript server bundle

1 MiB, one self-contained file

No dynamic imports. See JavaScript.

JavaScript server network allowlist

At most 32 hostnames and 16 secret names

Hostnames only, reached over https. See JavaScript.

JavaScript tool call

10 MiB memory, 10 seconds execution

Per tool call, including work still pending in promises. Each fetch gets 8 seconds and a 5 MiB response. Warm instances reconcile against stored configuration every 10 seconds. See JavaScript.

OpenAPI inline spec

3 MiB

Host larger specs and pass the URL. The server truncates tool names longer than 64 characters. See OpenAPI.

SharePoint file tools

Upload 4 MiB of UTF-8 text, download 10 MiB

List tools return 100 items by default and accept a page_size up to 999. The search_content tool returns 25 hits by default and accepts a page_size up to 500. See SharePoint.

Metabase query results

25 MiB

Larger results return an error. Lower the query’s limit or pre-aggregate. See Metabase.

Zendesk ticket comments

Up to 500 inline

With include_comments=true on a single-ticket fetch. See Zendesk.

Blocked input message and blocked output message

1 to 500 characters each

Returned to the caller when a policy blocks a prompt or a response. See Create a Guardrail.

Denied topics

Up to 30 per guardrail

Topic name 1 to 100 characters, definition up to 1,000 characters, up to five example phrases of up to 100 characters each. See Guardrail Policy Reference.

Sensitive information custom patterns

Name 1 to 100 characters, RE2 pattern 1 to 500 characters

RE2 doesn’t support lookaround. See Guardrail Policy Reference.

Word filters

Up to 10,000 custom words in total, and up to 4 managed word lists

The word total counts every word across all groups, which is an Amazon Bedrock limit. See Guardrail Policy Reference.

Automated reasoning policies

1 or 2 per guardrail

Each is a versioned policy ARN. See Guardrail Policy Reference.

Cost unit

Microcents: 1 cent = 1,000,000 microcents, so $1 = 100,000,000 microcents

Divide the cost in microcents (totalCostMicrocents in API responses) by 100,000,000 for dollars. See Set Up Budgets.

Budget periods

Daily, weekly, or monthly, calendar-aligned in UTC

Daily resets at 00:00 UTC, weekly at 00:00 UTC Monday, monthly at 00:00 UTC on the first of the month. See Set Up Budgets.

Budgets per tenant

One default plus at most one override per agent

The override’s target agent can’t change after creation. See Set Up Budgets.

At the budget cap

HTTP 429 until the period resets

The warning threshold must be greater than zero and less than the limit. When spend crosses it, responses carry a SpendLimit-Warning header and still succeed. See Set Up Budgets.

Cost and usage date range

Default last 7 days

Also last 14 days, 30 days, 90 days, month to date, quarter to date, year to date, and a custom range in UTC. See Analyze Cost and Usage.

Spend breakdown rows

Up to 5,000 per grouping

The Spend breakdown table on the Cost and usage page loads at most 5,000 rows for one grouping. When the window holds more, the table says so and suggests a narrower range. See Analyze Cost and Usage.

Audit log time range

Last 15 minutes to last 14 days, or a custom range

Captured request and response bodies are cut at 16 KB. See Review the Audit Log.

Agent network window

Default last 24 hours

Also last hour, 7 days, 30 days, 90 days, last year, or a custom range. See Visualize Your Agent Network.

Spans per export request

10,000

Exceeding it fails the whole request. Reduce the exporter’s batch size. See Self-Managed Agent Telemetry Reference.

Export request payload

4 MiB

Exceeding it fails the whole request. See Self-Managed Agent Telemetry Reference.

Attributes per span

256

Enforced per span with an OTLP partial success. See Self-Managed Agent Telemetry Reference.

Single attribute value

4 MiB

Matches the export request payload, so the request limit is the one that binds in practice. Enforced per span with an OTLP partial success. See Self-Managed Agent Telemetry Reference.

Span timestamps

Within the last 30 days, up to 10 minutes in the future

Enforced per span. See Self-Managed Agent Telemetry Reference.

Validation log window

Last 24 hours

Shows permanent drops only, grouped by cause, on the agent’s Setup tab. See Validate ingestion.

Tools per agent

10 to 15

Agents with more than 20 to 25 tools often show degraded tool selection. Partition tools across subagents. See Choose an Agent Architecture.

Max iterations

10 to 20 for simple queries, 20 to 40 for balanced workflows, 40 to 100 for complex analysis

Design workflows to finish in 20 to 30 iterations. Each iteration consumes tokens and adds latency. See Create an Agent and Why iterations matter.

Tool latency

Under 2 seconds

Slow tools multiply across iterations. See Troubleshoot Agents.