OS AI Agent versions independently of the Collectord agent and the Monitoring apps, using standard semantic versioning.
1.1.1
- Security: Built with Go 1.27.3, which refuses server certificates issued by eight eMudhra intermediate certificate authorities revoked for a key compromise
- Set
context_windowon every provider to the context length its model is loaded with. The agent now starts summarizing older tool results as a conversation nears it, and sizes the default reply length from it on Anthropic providers and for Claude 4 and later models on Bedrock. Without it, providers on Anthropic’s own API and Claude models on Bedrock count as 200,000 tokens and every other provider as 32,768, so a model loaded with less can run out of context - Bug fix: Long investigations re-ran their own searches and could spend their budget without writing an answer, because older tool results were hidden from the model after two rounds. Tool results are now sent in full until the conversation nears the provider’s context window, then summarized with the values the agent used
- Bug fix: Long answers stopped mid-sentence on Claude Opus 5.5 and Sonnet 5.5 through the Anthropic provider, and on Claude 4 and later models on Bedrock, because thinking counts against the 4,096-token default reply length. The default is now 32,000 tokens, or a quarter of the provider’s context window when that is under 128,000 tokens, and
default_max_tokensstill overrides it; set it for a Bedrock application inference profile, which keeps Bedrock’s default. An answer cut off at that limit, at the model’s context window, or by the provider’srequest_timeoutwhile it is being written is kept, says which, and is recorded as truncated in the activity trail; one cut off by the timeout used to fail the turn - Bug fix: One request could make more than a thousand model calls over about an hour, because its limit of 128 calls started again each time a long request continued in the background. The limit now covers the whole request, with one call more only when the background time budget interrupts the last, and that last call writes an answer from what was found so far, saying it was cut short, instead of the request failing with none; an investigation stopped this way writes no decision event, as before
- Bug fix: In one investigation the agent could run the same search many times over, each time costing a search. A call identical to an earlier successful one in the same request now gets that result back without running again, the chat marks its card as a repeat, and after three rounds of nothing but repeats the agent is asked for its answer
- Bug fix: The Bedrock model list offered base model IDs that Bedrock refuses for current Claude models; it now offers their inference profiles, such as
global.andus.. The docs name the IAM permissions a Bedrock provider needs, includingbedrock:CallWithBearerTokenfor a Bedrock API key andbedrock:ListInferenceProfilesfor the model list - Bug fix: A date given to the agent, such as the “observed 2025-08-01 to 2025-08-14” of a threat report, was refused by Splunk as a search time bound, and the agent searched the last 24 hours instead. A date, a date-time with or without an offset, and Splunk’s MM/DD/YYYY now work as
earliestandlatest, in the user’s Splunk timezone, withlateston a date including that whole day - Bug fix: The agent took “now” in the search head’s timezone, usually UTC, while Splunk shows each user search results in their own timezone, so it misread the times in results and shifted its relative time ranges by the difference. It now uses the user’s Splunk timezone preference (the run-as user’s in an alert investigation or
| osai task=), and the search head’s when the user has set none
Requires Splunk Enterprise 9.4, 10.0, 10.2, 10.4, or 10.6 on a search head with KV Store, or Splunk Cloud. See Requirements.
1.1.0
- New: A license covers a number of users, the Splunk users holding
use_os_ai_agent(Splunk administrators are not counted), and no longer caps providers. The License tab shows the licensed users beside the users holding the capability, counted every 5 minutes; more users than the license covers is a warning on the License tab and a bulletin message, and nobody is refused - New: The Health dashboard and the health check report decision events Splunk dropped because their decision index does not exist or is disabled, which a run otherwise reports as written while the alert meant to act on it never fires; the alert action’s and
| osai’s docs say the index must exist first - Security: Built with Go 1.27.2, which fixes vulnerabilities in the Go standard library that scanners report against Go 1.27.1, in the HTTP/1 and HTTP/2 client and server, TLS and MIME header parsing
- Bug fix: On Splunk Free every page and
| osairefused with a capability error, and the health check logged an error every 5 minutes. Splunk Free’s admin, which Splunk reports with no capabilities, now holds the app’s capabilities, the health check skips with one INFO line, and| osai run_as=fails, saying Splunk Free has no token authentication; see Splunk Free in the app’s README - Bug fix: An Ask OS AI Agent alert whose investigation was refused (a budget, the active-job cap, the free tier’s seat) left an empty conversation in the run-as user’s Chat for every firing
- Bug fix: On Splunk 9.4 and 10.0 the agent could build a dashboard with a timeline panel, which those releases do not have, with no warning; the panel showed “splunk.timeline is not defined”. The approval card, the result and
verify_dashboardnow warn and give a stacked column chart that draws on every supported release - Bug fix: Alert investigations need token authentication on Splunk Cloud too: it is off by default on new Splunk Cloud stacks, as on Splunk Enterprise. The app’s docs said Splunk Cloud had it on, and the message a failed investigation shows named only Splunk Enterprise
Requires Splunk Enterprise 9.4, 10.0, 10.2, 10.4, or 10.6 on a search head with KV Store, or Splunk Cloud. See Requirements.
1.0.0
Initial release.
- Chat with per-action approval. A multi-turn agent with server-side conversations, run by a background worker. In Chat, every write, and every search that uses more than Splunk’s commands that only read, pauses on an approval card you approve, deny, or edit first. See Chat.
- Bring your own model. OpenAI-compatible endpoints, local models included (Ollama, LM Studio, vLLM, llama.cpp), Anthropic, Azure OpenAI, and AWS Bedrock, with retries, failover, and Splunk’s proxy settings. See Providers.
- Dashboards and alerts it builds. Dashboard Studio dashboards, and saved searches, reports, and alerts with any installed alert action, grounded in embedded copies of the SPL and Dashboard Studio references. See Tools.
- Unattended investigations. The “Ask OS AI Agent” alert action and
| osai task=investigate read-only with nobody watching, and can record a decision you define for an alert of yours to act on;| osaialso answers prompts and explains SPL in any search. See Alerts and schedules and The osai command. - Knows your deployment. It reads your macros, lookups, data models, and installed commands as the signed-in user, follows your query templates and guidance, and uses the add-on commands an administrator allows, such as Kubernetes Search or Search for AWS, for data that is in no index. See Configuration.
- Costs, limits, and an activity trail. Costs at the rates you enter, per-user and deployment-wide budgets, and one event per turn, tool call, refusal, and Setup change, read by the Token Usage, Activity Trail, and Health dashboards. See Costs and limits and Dashboards.
- Your permissions are the boundary. Every tool runs with the calling user’s own Splunk permissions, provider keys never reach a process a user controls, job credentials are encrypted at rest, and the app sends no telemetry. See Security.
- Search head clustering, an MCP server, and a free tier of one provider and one user for non-commercial use. See Licensing.
Requires Splunk Enterprise 9.4, 10.0, 10.2, 10.4, or 10.6 on a search head with KV Store, or Splunk Cloud. See Requirements.
Known issues and limitations
- Alert investigations need token authentication, which is off by default on Splunk Enterprise.
- Right after an app upgrade without a splunkd restart,
| osaimay refuse to run; restart splunkd, or wait for the handler processes to recycle. - On Splunk 10.6, the KV Store logs stored values into
_internalat its default log level, conversations included; see Security. - Azure managed identity needs an Azure-hosted search head; it does not apply on Splunk Cloud.