Stay up to date with the latest features, improvements, and updates to the Kaman AI platform.
v3.28
Relevance Lab Makes a Strategic Majority Investment in kaman.ai
September 11, 2026
Relevance Lab has made a strategic majority investment in kaman.ai
A news banner at the top of the kaman.ai website links to the full announcement
Website
The public kaman.ai pages, including the homepage, pricing, blog, papers and documentation, now show a news banner above the navigation announcing Relevance Lab's strategic majority investment. Selecting the banner opens the announcement on the Relevance Lab newsroom in a new tab.
The banner appears only on the public website. Signed-in workspace screens such as the dashboard, chat and editors are unchanged.
Previous Releases
v3.27
September 1, 2026
Data Lake Stability: No More Interruptions During Storage Cleanup
Fixed repeated data lake restarts that briefly interrupted queries, dashboards, and scheduled workflows
Routine storage cleanup no longer runs while queries are reading the same data
Housekeeping now runs once a day in a predictable window instead of every half hour
Fixes
The data lake service could restart every few minutes on workspaces with continuous data synchronization. Its routine housekeeping removes retired storage files in the background, and this could run at the same moment a query was reading from that same storage, stopping the service. Housekeeping now waits for in-progress queries to finish and holds them briefly while it runs, so the service stays up.
Storage housekeeping used to start every 30 minutes. On large workspaces a single pass takes over ten minutes to review all stored files, so cleanup was running almost continuously. It now runs at most once a day, and the schedule is configurable per deployment.
Triggering maintenance manually from the admin API now takes the same safeguards as the scheduled run, so it can no longer interrupt live queries.
v3.26
August 21, 2026
Platform Stability: Faster Recovery From Heavy Data Sync Load
Resolved brief platform-wide interruptions that occurred every few minutes during periods of very heavy data synchronization
Live data syncs, scheduled jobs, and chat sessions are no longer affected when the internal cache fills up
The platform now automatically makes room in its cache instead of restarting under memory pressure
Fixes
During sustained high-volume data synchronization, an internal cache could fill to its limit and restart itself every few minutes. Each restart briefly interrupted scheduled jobs, live data updates, and in-progress conversations. The cache now automatically clears its oldest entries instead of restarting, so heavy sync activity no longer causes interruptions.
Background AI enrichment of synced records now keeps its working data for a shorter, right-sized period, significantly reducing memory usage during large synchronizations without affecting the quality or completeness of the enrichment.
v3.25
August 20, 2026
Tool Search Results No Longer Disappear While Building an Agent
Searching for a workflow or plugin in the Select Resources dialog now reliably shows every match
Results no longer vanish from the list after you tick a checkbox or click a resource
Workspaces with a large number of tools now see a stable, alphabetically ordered resource list
Fixes
In the agent editor's Select Resources dialog, matching workflows and plugins could disappear from the list even though the search had found them. Selecting an item quietly reloaded the unfiltered resource list behind the scenes, and your search was then applied to that shorter list instead of the actual search results. Search results now stay on screen for the full time the dialog is open.
Workspaces with more than 100 plugins and workflows were showing a different arbitrary subset of resources on each visit, because the list was cut off without a defined order. The list is now sorted alphabetically, so the same resources appear consistently every time.
Selecting a resource no longer collapses the expanded details you were viewing in the dialog.
v3.24
August 3, 2026
Saved Tool Settings Now Apply Every Time a Tool Is Used
API keys and other saved tool settings are now used even when the assistant finds the tool on its own
Fixes tools failing with "missing key" style errors despite the key being saved
Affects every tool with user-configured fields, including scheduled and automated runs
Fixes
Tools that store your own settings — an API key, an account ID, a workspace name — now receive those values in every situation. Previously they were applied only when the tool was assigned to the agent up front; if the assistant discovered the tool by searching for it mid-conversation, the saved values were dropped and the tool ran without them.
This typically showed up as the external service rejecting the request for a missing key or token, even though the key was saved correctly in your settings — with no indication that anything was wrong with the configuration.
Scheduled and automated agent runs are affected the most, since they frequently discover tools on demand. Those runs now pick up saved tool settings the same way an interactive chat does.
v3.23
July 25, 2026
Long Conversations Stay On Track
Your question is never lost when a long chat is condensed mid-answer
The assistant no longer restarts and re-introduces itself in the middle of a task
Earlier parts of a conversation are kept even if the summarizing step is unavailable
Fixes
When a conversation grows long enough to be condensed automatically, your most recent request is now always kept word-for-word. Previously, during a long run of tool steps, the request could be folded into the summary — leaving the assistant with nothing to answer, so it would greet you as if the chat had just started.
If the summarizing step times out or is unavailable, the older part of the conversation is now preserved as-is instead of being discarded. Earlier, that content could be dropped entirely, silently losing context you had already provided.
The condensing step no longer inherits the assistant's own persona, which occasionally caused it to reply in character ("Hi, I'm …, how can I help?") and have that greeting saved in place of your conversation history.
Very long conversations are now trimmed more carefully so the assistant is never left with only its instructions and no chat, and so tool results are never separated from the steps that produced them.
v3.22
July 13, 2026
Smoother Sign-In for People Who Belong to More Than One Organization
Invited teammates can now join a new organization even if they already have an account elsewhere on Kaman
Verification emails now reach the right account when the same email exists in multiple organizations
Clearer guidance when an account is waiting on an invitation instead of email verification
Fixes
Accepting an organization invitation no longer fails with "Username is already taken" when you already use the same name in another organization — names only need to be unique within each organization.
Resending a verification email now targets the account that actually needs verifying, instead of stopping early because a copy of your email in another organization was already verified.
If your account is still waiting on an organization invitation, the resend-verification screen now tells you to use the invitation link (or ask your admin to resend it) rather than sending a verification email that couldn't complete your setup.
v3.21
July 6, 2026
Lock Agents Down to Only the Tools You Assign Them
New "Restrict to Assigned Tools" switch when creating or editing an agent
A restricted agent can only use built-in system tools and the tools you explicitly give it
Closes the gaps that let agents discover and use tools outside their assigned set
New Features
Agent settings now include a "Restrict to Assigned Tools" switch. Turn it on and the agent is locked to its assigned tools plus Kaman's built-in system tools — it can never find or use anything else in your workspace, even if you assign it no tools at all.
Restricted agents no longer receive marketplace tool suggestions, so nothing outside their approved toolset is ever offered to them.
Fixes
Fixed several gaps where an agent could reach tools beyond the ones configured for it — including tool lookups that matched by name and could land on a similarly-named tool from a different extension.
Sub-agents spawned by a restricted agent now inherit the same tool restriction automatically.
v3.20
June 23, 2026
Safely Recover Data Sources Without Losing What's Already There
Re-syncing a data source now asks whether to add to your data or fully replace it
A new additive sync backfills missing records without dropping your existing rows
The safe choice is the default — full replace is now an explicit, deliberate action
New Features
When you sync a data source, you can now choose Additive — it pulls the latest data and merges it in, keeping every existing row. This is the safe way to recover or backfill records that went missing without risking the rest of your data.
Full replace (clear the table and reload everything) is still available, but it's now a clearly-labeled, intentional choice rather than the automatic behavior.
Improvements
The Sync button now opens a short prompt explaining each option, so it's always clear whether a sync will add to your data or overwrite it.
v3.19
June 10, 2026
Run Workflows in the Background and Track Their Progress
Kick off a workflow and get an immediate run ID instead of waiting for it to finish
Check a running workflow's live status and see exactly which step it's on
Workflows that pause for your input are now clearly flagged as waiting, not failed
New Features
Workflows can now be started in the background — you get a run ID right away and can keep working while the workflow runs.
A new status check lets you see whether a run is still going, finished, failed, or waiting on you, along with the step it's currently working on.
Once a background workflow finishes, its full result is available from the same status check — so you get the output back without re-running anything.
Improvements
When a workflow pauses because it needs information from you, it's now shown as "waiting for input" with the exact fields required — instead of looking like an error.
v3.18
June 7, 2026
Agents Recover Gracefully Instead of Stopping Silently
No more "agent stopped unexpectedly" with no explanation — temporary glitches now retry automatically
When something genuinely can't be answered, you get a clear, specific reason instead of a blank stop
Smarter automatic recovery from busy services, timeouts, and empty model responses
Improvements
The assistant now automatically retries when the AI service is briefly busy, times out, or returns an empty response — so short-lived hiccups resolve themselves without you having to resend.
Each kind of problem now gets the right response: transient issues are retried, overly long conversations are trimmed and re-tried automatically, and only genuinely unrecoverable issues are shown to you.
Fixes
Fixed the case where the assistant could stop with "The agent stopped unexpectedly without producing a response" and no detail. These situations are now either recovered automatically or explained with a clear, actionable message (for example, a content-safety block or a service being temporarily unavailable).
v3.17
May 24, 2026
No More Stuck Agents on Long Conversations
Agents on long chats no longer appear to hang while quietly re-summarizing in a loop
Faster responses on heavy workflows — no wasted summarization passes between turns
Same aggressive context management, without the silent loop
Fixes
Fixed an issue where, after a conversation grew past the compression threshold, the agent could re-summarize the same context every turn without making progress — making it look like the agent was stuck. Compression now skips a pass when nothing meaningful has been added since the last one.
v3.16
May 19, 2026
More Breathing Room for Long-Running Agents
Agents working on large artifacts no longer time out mid-response
Complex multi-step tasks have far more room before hitting depth limits
Smoother experience for heavy workflows like reports, code generation, and data tasks
Improvements
Streaming chunk timeout raised from 1 minute to 10 minutes — large artifact generation no longer cut short when the model takes longer to respond
Per-message iteration ceilings raised 10× — long multi-tool conversations and intricate workflows no longer hit the safety brake on legitimate work
Agent graph depth limit raised from 50 to 500 — the “maximum processing depth” error during long multi-step tasks should now be far less common
Sub-agent iteration default raised from 25 to 250 — delegated sub-tasks have more room to finish without manual override
WebSocket chat depth limit raised from 100 to 1000 — chat-based agents match the rest of the platform
Model proxy request timeout raised from 2 minutes to 10 minutes — long Claude/OpenAI/Bedrock streaming responses no longer silently end mid-stream on large-context turns
v3.15
April 29, 2026
Smarter Recovery, Tighter Boundaries
Streaming responses recover gracefully on provider hiccups — no more half-rendered messages
Model fallback now kicks in on any provider issue, not just rate limits
Memory and connections cleaned up automatically across long-running processes
Improvements
Streaming responses end cleanly with a proper completion marker even when the upstream model errors mid-stream
Model fallback chain now retries on server errors, network failures, and provider timeouts — not just rate limits
VM tool latency reduced — agents now share a single response channel instead of opening one per call
Chat history and iteration-tracking memory now self-evict idle entries so long-running processes don't grow unbounded
Fixes
Personal VM secret comparison hardened against timing-attack inference
Git webhooks refuse unsigned requests instead of falling through silently when no secret is configured
VM debug proxy restricted to admins and private IP ranges only — no longer a relay for arbitrary traffic
v3.14
April 29, 2026
Sturdier Conversations, Cleaner Failures
Agents no longer silently forget mid-conversation when the database hiccups
Stuck LLM responses now fail fast instead of hanging until the request timeout
Tools are quicker to respond — no more 60-second waits for replies that already arrived
Improvements
Conversation memory now recovers automatically from temporary database issues
Detect hung LLM responses within 60 seconds — fall back instead of waiting minutes
Tool execution responses are picked up immediately even when they arrive instantly
Fixes
Closed a path where two clients could collide on the same session ID and abort each other's stream
Background history writes can no longer fire after a session ends
Health endpoint now reports conversation-state persistence outages as degraded
v3.13
April 28, 2026
Tighter Permissions, Safer Billing
Permission grants are now strictly scoped to your organization
Billing flows now use your authenticated identity — no more impersonation risk
Internal service authentication hardened across the platform
Improvements
Resource permission management is now restricted to the org owner or org admins — no more cross-tenant edits possible
Newsletter sending is now admin-only
Build pipeline made resilient to transient DNS issues during dependency fetch
Fixes
Closed a path where billing checkout could be initiated under another user's identity
Removed hard-coded service-to-service fallback secrets that could ship to production unintentionally
Webhook authentication now refuses to operate when its secret is unconfigured
v3.12
April 27, 2026
Faster, Steadier, Easier to Diagnose
Agent answers feel more consistent — same question, same approach
Faster first responses on simple chats (less waiting before the agent starts typing)
When something goes wrong, you can finally see exactly what happened in your session
New Features
Per-session log viewer — paste a session ID, get the full timeline across services in one view (for ops and support)
Health dashboard now shows live status of every backend dependency (database, cache, queue, model proxy) at a glance
Improvements
Agents now respond faster on simple text chats — skipping work that wasn't needed
Agent answers are more consistent run-to-run — same question gets the same approach
Plugin and workflow saves return to you immediately, even if the background sync is busy
Auth checks are now cached for a short window — fewer round-trips, faster API responses
Database connection pooling is loud about real outages instead of pretending things are slow
Fixes
Fixed a data-loss path where buffered writes could vanish if the database hiccupped at exactly the wrong moment
Closed gaps where row-level security predicates could silently degrade under malformed inputs
Removed a path where encrypted credentials could fall back to being treated as plaintext
Stopped a bug where DSL imports could create duplicate tools/agents on transient database errors
Scheduled tasks no longer get stuck in a 'still running' state when their notification step fails
Agent Hub real-time updates now log failures instead of silently dropping them
v3.11
March 25, 2026
Free Monthly Credits, Agent Delegation, and Trust & Transparency
Every new account now gets free credits each month — no credit card required
Agents can now delegate tasks to other agents for complex multi-step work
Configurable credit display, improved billing, and updated legal pages for full transparency
New Features
Free monthly credits for all new accounts — explore agents and workflows at no cost
Configurable credit display multiplier — credits are now shown as display credits (e.g. 5,000 credits instead of $5)
Agent delegation — agents can now call other agents to break down complex tasks across specialized roles
Delegate agent UI — teal-accented progress cards and modal show when an agent hands off work to another
Audit Trail section added to Security Fabric compliance page