What kagent-dev/kagent shipped
Written by FoxPlug from the repository's own releases; not affiliated with Kagent. Paste your own repository in the box below and FoxPlug writes the same weekly update for your product. This page is an automatic summary of the releases, pull requests and commits in github.com/kagent-dev/kagent; Kagent did not write it and does not use or endorse FoxPlug. Every line links to the change it describes.
Get a weekly update like this for your product, free
Week of November 2, 2026
What shipped
- The core MCP server gained session lifecycle and read-back tools, including starting, deleting, renaming, and reading back what an agent did. Pull request #3339
- Added UI management for EgressPolicies with a Policies page, search, namespace filter, a Used by table, and create and edit forms that check the CRD rules before saving. Pull request #3542
- MCP clients can now discover agents with
list_agentsand start a conversation with an agent directly. Pull request #3486 - Runtimes now receive compiled config and Agent Cards through immutable Postgres-backed bundles instead of ActorTemplate env vars. Pull request #3515
- Session deletion now returns an error while turns are active and offers an explicit force option across the API, CLI, and UI. Pull request #3530
- Added RevertSession so a crashed runtime can be reverted to its last durable snapshot when a client asks. Pull request #3395
- Runtimes now record the gateway-forwarded user as user.id instead of enduser.id on the request span and its child spans. Pull request #3552
- ModelConfig conditions now move their transition time only on a status change, matching the API conventions used by agent conditions. Pull request #3340
- Restored upgrade tests to select the new release candidate and check for the optional tracking table and Harness fixtures. Pull request #3558
Why it matters
This week adds self-hosted sandbox support, fuller MCP session tools, and a UI for managing egress policies, which gives developers more control over how agents run and how they are discovered and driven. The runtime configuration moves to durable bundles, and session deletion and crash recovery gain explicit options, so behaviour around sessions is clearer. Several fixes address status handling, task finalization, and upgrade tests, which reduce failures during normal use and upgrades.
Changelog entry
- Add session lifecycle and read-back tools to the core MCP server. Pull request #3339
- Add a Policies page with an Egress Policies tab and create and edit forms in the UI. Pull request #3542
- Add
list_agentsand renameinvoke_sessiontoinvoke_agentfor MCP clients. Pull request #3486 - Deliver runtime config and Agent Cards through immutable Postgres-backed bundles. Pull request #3515
- Return an error on session deletion while turns are active and add a force option. Pull request #3530
- Add RevertSession to recover crashed runtimes to their last durable snapshot. Pull request #3395
- Record the gateway-forwarded user as user.id on request spans and their children. Pull request #3552
- Move a ModelConfig condition transition time only on a status change. Pull request #3340
- Finalize failed tasks after execution cancellation so the runtime releases its worker. Pull request #3541
- Retry TaskStore calls through a controller restart so turns do not stay WORKING. Pull request #3391
- Restore upgrade tests from the new release candidate and check optional tracking and Harness fixtures. Pull request #3558
This week: self-hosted sandbox support, session lifecycle and read-back tools in the MCP server, agent discovery, egress policy management in the UI, and config delivered through durable bundles.
Week of October 26, 2026
What shipped
- The first release candidate, v1.0.0-rc1, is published, bundling improved chat TUI rendering and input, a sandbox guest service redesign, and outputSchema formatting, pattern and type lists with opt-in strict mode. Release
- ModelConfig and RemoteMCPServer now accept an actorJwt credential, and EgressPolicy authentication supports it too, so downsteam services can identify the executing Actor at the gateway. Pull request #3359
- Authentication now gives human, service-account and Actor callers a shared opaque Principal.ID in place of the separate user and agent identity fields. Pull request #3358
- AgentTemplates and inline Agent templates can reference ModelConfigs and RemoteMCPServers in other namespaces, checked against the target's allowedNamespaces at compile time. Pull request #3382
- A new kagent install --dry-run writes a reviewable Helm installation recipe without touching the cluster, and the exporter is exposed as install.Export. Pull request #3432
- The controller now retries golden builds that failed for infrastructure reasons, so an Agent or SandboxTemplate is no longer stuck after a single error. Pull request #3335
- Installation no longer passes model API keys in Helm --set arguments, carrying them instead as secret values written to a 0600 file. Pull request #3446
- Install now honors the configured controller service account instead of granting Substrate access and PostgreSQL certificates only to the hardcoded kagent-controller account. Pull request #3406
- A2A push callbacks are delivered concurrently with config.push.maxConcurrency and config.push.sendTimeout, defaulting to 10 in-flight deliveries. Pull request #3366
- The Helm chart adds sandboxConfigRef under substrateWorkerPool so a user-defined SandboxConfig can override the default. Pull request #3357
Why it matters
This week carries the project to its first release candidate while closing out pre-RC API cleanup that would have been costly to do later. Users get cross-namespace ModelConfig and RemoteMCPServer references, install dry-run for reviewing Helm changes, and safer credential handling during installation. Alongside that, retries for failed golden builds and concurrent A2A callbacks reduce the ways a single fault stalls an Agent or delays other callbacks.
Changelog entry
- Release v1.0.0-rc1. Release
- Add actorJwt credentials to ModelConfig, RemoteMCPServer and EgressPolicy authentication. Pull request #3359
- Use a single Principal.ID for human, service-account and Actor callers. Pull request #3358
- Allow cross-namespace ModelConfig and RemoteMCPServer references checked against allowedNamespaces. Pull request #3382
- Add kagent install --dry-run to export a Helm installation recipe. Pull request #3432
- Extract installation and uninstallation into a reusable library. Pull request #3427
- Retry golden builds for Agents and SandboxTemplates after infrastructure failures. Pull request #3335
- Keep model credentials out of process arguments during install. Pull request #3446
- Honor the configured controller service account when granting Substrate access and certificates. Pull request #3406
- Deliver A2A push notifications concurrently with maxConcurrency and sendTimeout settings. Pull request #3366
- Add sandboxConfigRef under substrateWorkerPool in the Helm chart. Pull request #3357
- Remove unused Harness status and ADK network,
share_toolsand ADKNetworkConfig fields. Pull request #3454
V1.0.0-rc1 is out. This week also brought cross-namespace ModelConfig and RemoteMCPServer refs, kagent install --dry-run, retried golden builds, concurrent A2A callbacks, and model API keys kept out of process arguments.
Kagent v1.0.0-rc1 is published, the first release candidate, and the week's changes close out pre-RC API cleanup. AgentTemplates and inline templates can now reference ModelConfigs and RemoteMCPServers in other namespaces through allowedNamespaces. Installation gains a dry-run that writes a reviewable Helm recipe without touching the cluster, stops passing model API keys in Helm --set arguments, and grants Substrate access to the controller service account you configure. The controller retries golden builds that failed for infrastructure reasons, and A2A push callbacks are delivered concurrently with configurable limits and timeouts.
Week of October 19, 2026
What shipped
- Owned-resource RBAC is now enforced alongside ownership checks for Sessions, Checkpoints, Memory, Sandboxes and ScheduledRuns, and shared reads require recipient Session permission. Pull request #3299
- Substrate actor calls to the control plane are now authenticated with Substrate-issued JWTs in the Authorization header, and invalid or stale credentials no longer fall back to unauthenticated access. Pull request #3282
- Substrate reads and page tokens are now scoped to the watched namespaces so one install no longer sees another install's actors, actor templates or namespaces. Pull request #3261
- Built-in filesystem and bash tools are now available independently of skills, with Go and Python execution aligned to Substrate's sandbox and durable workspace. Pull request #3283
- PDF and Office file contents are now delivered end-to-end to Responses API models, and the chat UI accepts those files. Pull request #3200
- Kagent uninstall now removes everything kagent install creates and exits non-zero when a step fails. Pull request #3285
- The UI now keeps following a chat turn when its stream breaks instead of reporting the turn as failed while the task is still running. Pull request #3198
- AgentTemplate outputSchema now accepts format, pattern and type lists such as string or null, with an opt-in outputSchemaStrict setting. Pull request #3284
- A controller-wide runtimeDefaults.builtinTools.disableAskUser setting lets ADK agents run without the built-in
ask_usertool. Pull request #3298 - The kagent CLI chat experience has improved transcript rendering and input. Pull request #3292
Why it matters
This week the access-control and namespace-scoping work landed: resource permissions now apply across core resources, actor calls are authenticated, and Substrate reads stay within watched namespaces. It also brings practical fixes for people running agents, including uninstall that removes everything install creates, file support end-to-end for PDFs and Office files, and a chat UI that keeps following a turn when the stream breaks.
Changelog entry
- Enforce owned-resource RBAC and scoped Substrate inventory for Sessions, Checkpoints, Memory, Sandboxes and ScheduledRuns. Pull request #3299
- Authenticate Substrate actor calls to the control plane with Substrate-issued JWTs. Pull request #3282
- Scope Substrate reads and page tokens to the watched namespaces. Pull request #3261
- Enforce scoped authorization for Kubernetes-backed resources across CRUD and custom list services. Pull request #3281
- Enable built-in filesystem and bash tools independently of skills and simplify sandbox execution. Pull request #3283
- Send PDF and Office file contents to Responses API models end-to-end and accept them in the chat UI. Pull request #3200
- Make kagent uninstall remove everything kagent install creates and exit non-zero on failure. Pull request #3285
- Keep following a chat turn in the UI when its stream breaks while the task is still running. Pull request #3198
- Allow format, pattern and type lists in outputSchema and add opt-in outputSchemaStrict. Pull request #3284
- Add config.runtimeDefaults.builtinTools.disableAskUser to disable the
ask_usertool. Pull request #3298 - Apply the Session idle TTL to settled INPUT_REQUIRED and AUTH_REQUIRED turns. Pull request #3254
- Improve chat TUI rendering and input in the kagent CLI. Pull request #3292
This week: owned-resource RBAC and authenticated actor calls, Substrate reads scoped to watched namespaces, filesystem and bash tools without skills, PDF and Office files end-to-end, and an uninstall that removes everything install creates.
This week in kagent: resource permissions now apply alongside ownership checks across Sessions, Checkpoints, Memory, Sandboxes and ScheduledRuns. Substrate actor calls are authenticated with Substrate-issued JWTs, and Substrate reads and page tokens stay within the watched namespaces. Built-in filesystem and bash tools now work independently of skills, PDF and Office files reach Responses API models end-to-end, and kagent uninstall removes everything install creates. The chat UI also keeps following a turn when its stream breaks.
Week of October 12, 2026
What shipped
- Released v1.0.0-alpha14, which includes built-in Harness images defaulted from release charts and runtime suspension migrated to DATA snapshots. Release
- Added a token passthrough option on RemoteMCPServer that forwards the incoming bearer token as the server Authorization header. Pull request #3257
- Added reusable EgressPolicy resources that agents and sandbox templates select through spec.egressPolicyRefs, with header and Secret-based authentication. Pull request #3177
- Changed CLI agent and agent template reads to go through the kagent API instead of Kubernetes, so the CLI can work without a kubeconfig when an API URL is given. Pull request #3232
- A2A now reports task token usage through a versioned extension, with both executors delegating to the library A2A executor. Pull request #2461
- OCI skills now mount through Substrate image volumes, moving registry pulls out of the runtime and using the same registry credentials as workload images. Pull request #3203
- Added runtime.skill_tools.disable_bash to turn off the built-in bash tool for the root agent and subagents while keeping skill loading and file tools. Pull request #3222
- Made session share tokens read-only, so shared conversations no longer allow replies, approvals, cancellations or retries. Pull request #3189
- Expired and missing A2A sessions now return task not found for both context and task identifiers, including missing agents. Pull request #3231
- Failed checkpoints can now be deleted, so they no longer keep a runtime revision pinned. Pull request #3199
Why it matters
This week brought the v1.0.0-alpha14 release along with changes that affect how agents reach the network, how the CLI talks to the API, and how shared sessions behave. Several fixes improve error reporting and cleanup, such as consistent not-found responses for A2A sessions and deletion of failed checkpoints. The changes to token passthrough, egress policies and read-only share tokens are worth reviewing before upgrading.
Changelog entry
- Released v1.0.0-alpha14. Release
- Added token passthrough for remote MCP servers. Pull request #3257
- Added reusable egress policies for agents and sandboxes. Pull request #3177
- Read agents and agent templates through the API instead of Kubernetes in the CLI. Pull request #3232
- Reported task token usage through a versioned A2A extension. Pull request #2461
- Mounted OCI skills through Substrate image volumes. Pull request #3203
- Allowed disabling the built-in bash tool in the Go ADK. Pull request #3222
- Made session share tokens read-only. Pull request #3189
- Returned not found for expired and missing A2A sessions. Pull request #3231
- Allowed cleanup of failed checkpoints. Pull request #3199
- Told UpstreamAuth which A2A method it is authorizing. Pull request #3178
- Stopped building the Python ADK image. Pull request #3221
Kagent v1.0.0-alpha14 is out. This week: RemoteMCPServer token passthrough, reusable EgressPolicy resources, CLI reads through the kagent API, read-only session share tokens, and A2A token usage reporting. Failed checkpoints can now be deleted.
Kagent v1.0.0-alpha14 is released. This week's changes include an opt-in token passthrough for RemoteMCPServer, reusable EgressPolicy resources selected by agents and sandbox templates, and CLI agent reads that go through the kagent API so a kubeconfig is not required when an API URL is given. Session share tokens are now read-only, expired and missing A2A sessions return task not found consistently, and failed checkpoints can be deleted instead of pinning a runtime revision. OCI skills now mount through Substrate image volumes, and A2A reports task token usage through a versioned extension.
Week of October 5, 2026
What shipped
- Released v1.0.0-alpha11 with a configurable database image plus fixes for unit tests, scheduled run timeouts and flaky browser tests. Release
- Added A2A task push notifications, with callbacks accepted on sends and task callback management through A2A, the CLI and the UI. Pull request #3036
- Read_sandbox_file now returns text as numbered, paged lines and images directly, instead of raw base64. Pull request #3120
- Moved Kagent tables into a kagent schema owned by
kagent_owner, with a configurable pgvector schema. Pull request #3111 - Made the bundled PostgreSQL image used by kagent install configurable through KAGENT_BUNDLED_POSTGRES_IMAGE, for mirrors and air-gapped clusters. Pull request #3112
- Made kagent install exit non-zero when a step fails, so scripts and Make targets no longer continue after a failed install. Pull request #3109
- Session quiescence workers are now woken after task settlement commits, keeping pause and suspend prompt while idle recovery scans run about once a minute. Pull request #3074
- Added a Helm value to set allowedNamespaces on the built-in tool server, so agents in other namespaces can reference it without duplicating the server. Pull request #3145
- The chart now runs Substrate worker pods under their own ServiceAccount instead of the default one. Pull request #3067
- Isolated worker pool placement by namespace so ActorTemplates cannot select another namespace's workers with the same pool name. Pull request #3089
Why it matters
This week lands another alpha with configuration and install fixes that matter for real deployments: the database image is now configurable, Kagent tables move to their own schema, and failed installs exit non-zero instead of silently succeeding. A2A task push notifications and better sandbox file reads expand what agents and clients can do, while several fixes keep multi-namespace and shared-cluster setups from leaking across boundaries.
Changelog entry
- Release v1.0.0-alpha11. Release
- A2A task push notifications with callback management. Pull request #3036
- Read_sandbox_file returns text pages and images. Pull request #3120
- Kagent tables stored in a kagent schema with a configurable pgvector schema. Pull request #3111
- Database image installed by kagent install is configurable. Pull request #3112
- Kagent install exits non-zero on failure. Pull request #3109
- Session quiescence workers wake after task settlement. Pull request #3074
- Helm value for allowedNamespaces on the built-in tool server. Pull request #3145
- Substrate workers run under their own ServiceAccount. Pull request #3067
- Worker pool placement isolated by namespace. Pull request #3089
- Kagent install relies on Helm for install namespaces and rejects unmanaged CRDs. Pull request #3137
V1.0.0-alpha11 is out. This week: A2A task push notifications, sandbox files read as paged text and images, Kagent tables in their own schema, a configurable Postgres image for installs, and worker pools isolated by namespace.
V1.0.0-alpha11 shipped this week. A2A task push notifications are in, including callbacks on sends and task callback management through A2A, the CLI and the UI. Sandbox file reads now return text as numbered, paged lines and images directly rather than base64. On the install side, Kagent tables moved into their own schema, the bundled PostgreSQL image is configurable for mirrors and air-gapped clusters, and kagent install now exits non-zero when a step fails. Worker pools are isolated by namespace, Helm can set allowedNamespaces on the built-in tool server, and Substrate workers run under their own ServiceAccount.
Week of September 28, 2026
What shipped
- Made database polling intervals configurable for sandbox expiration, scheduled runs, and runtime cleanup to reduce serverless database activity. Pull request #3038
- Added expiration support to session shares so tokens can have a limited lifetime. Pull request #2987
- Replaced Kubernetes transaction-based helper names with CRUD verbs for clearer database operation semantics. Pull request #2859
- Fixed Helm PostgreSQL pool configuration to propagate to database connections. Pull request #3032
- Updated README documentation to reflect v1 architecture with AgentTemplates, Harnesses, Sessions, and PostgreSQL. Pull request #2996
- Added local AGENTS.md guides for protobufs, gRPC transport, PostgreSQL, and UI directories. Pull request #3035
- Repaired image scans and switched to uv-managed Python with production dependencies from uv.lock. Pull request #2995
- Pinned UI base image digest and apk package versions to major versions for build stability. Pull request #2951
Why it matters
V1.0.0-alpha7 brings model streaming control, better Helm configuration flexibility, and session share expiration for production use cases. Database polling configurability helps reduce costs on serverless deployments. Documentation and code quality improvements make the project easier to understand and extend.
Changelog entry
- Feat(helm): make the controller ServiceAccount optional Release
- Fix: allow disabling kagent model streaming Pull request #3028
- Feat(session): let a session share expire Pull request #2987
- Feat: configure database polling and skip unchanged MCP writes Pull request #3038
- Fix: propagate helm postgres pool variables to db connection Pull request #3032
- Docs: update README for the v1 architecture Pull request #2996
- Fix: repair image scans and use uv-managed Python Pull request #2995
- Refactor(core): replace update helpers with CRUD verbs Pull request #2859
- Docs: add local AGENTS.md guides Pull request #3035
- K: pin the UI base image digest and apk major versions Pull request #2951
Week of September 21, 2026
What shipped
- V1.0.0-alpha5 released with explicit Agent definitions, Sessions, SandboxTemplate and Sandbox contracts, and UI chat part renderer overrides. Release
- Agents are now explicit resources that pair AgentTemplates with Harnesses, replacing implicit label matching. Pull request #2952
- Standalone sandboxes let users and agents create scratch environments to run commands, move files, and manage lifecycle independently. Pull request #2962
- A2A execution moved from the gateway into agent runtimes with a private gRPC TaskStore for persistence. Pull request #2939
- Sandbox workflows and agent commands are now grouped under kagent sandbox for discoverable template and lifecycle management. Pull request #2966
- Environment settings standardized under KAGENT_ prefix with 138 documented variables generated into docs/env.md. Pull request #2967
- Telemetry exports semconv metrics through Prometheus and OTLP, drops ADK legacy spans, and correlates logs. Pull request #2929
- CI refactored to build release artifacts in a separate job before the final release job, improving parallelization. Pull request #2972
- Helm authentication settings wired into controller startup to enforce oauth2-proxy on controller requests. Pull request #2971
Why it matters
Alpha5 introduces explicit Agents as first-class resources and standalone Sandboxes as independent execution environments. A2A execution moved to runtimes enables better fault isolation and persistence. These changes prepare the project for stable APIs and production deployment patterns.
Changelog entry
- Feat(api): add SandboxTemplate and Sandbox contracts for standalone sandbox environments Release
- Feat: add explicit Agent definitions pairing AgentTemplates with Harnesses Pull request #2952
- Feat(sandbox): implement standalone sandbox runtimes with command execution and file transfer Pull request #2962
- Feat: move A2A execution into runtimes with a gRPC TaskStore for persistence Pull request #2939
- Feat: add sandbox workflows and group agent commands under kagent sandbox Pull request #2966
- Refactor: standardize environment settings under KAGENT_ prefix with generated reference Pull request #2967
- Feat(telemetry): export semconv metrics, drop ADK legacy spans, and correlate logs Pull request #2929
- Feat(ui): create, edit and delete explicit Agents through full-page forms Pull request #2953
- Fix: wire Helm authentication settings into controller startup Pull request #2971
- Fix(byo): centralize ADK runtime configuration for durable session URLs Pull request #2965
- Fix: flush buffered OTEL logs on agent checkpoint Pull request #2772
- Ci: overlap slow E2E harness scenarios to run in parallel Pull request #2969
V1.0.0-alpha5: explicit Agents pair templates with harnesses, standalone Sandboxes run independently, and A2A execution moves into runtimes. Standardized environment variables and improved telemetry.
Kagent v1.0.0-alpha5 ships explicit Agent resources and standalone Sandboxes. Agents now pair AgentTemplates with Harnesses explicitly instead of through label matching. Sandboxes run independently as scratch environments for users and agents. A2A task execution moved from the gateway into agent runtimes with private persistence. Environment variables standardized across all components with 138 documented settings. Telemetry improved with semconv metrics export and better log correlation.
Week of September 14, 2026
What shipped
- V1.0.0-alpha1 released with agent namespace in list view, controller affinity configuration, responses API support in harness, and topology improvements. Release
- Model and MCP secrets now use Substrate gateway credential injection instead of literal environment variables across all compilers. Pull request #2868
- Substrate actors and workers now load server-side with pagination, sorting, and search for clusters too large for single inventory reads. Pull request #2771
- Actor egress policies now provisioned with Substrate v0.0.30 to prevent HTTP 403 failures on model and MCP requests. Pull request #2839
- Go runtime now honors context compaction settings from the Agent CRD spec, matching Python runtime behavior. Pull request #2790
- AgentInstance lifecycle execution serialized to prevent concurrent retries from issuing duplicate mutations. Pull request #2857
- Snapshots now have names, and snapshot marks in chat carry Fork, Rename, and Delete actions. Pull request #2847
- Scoped authorization added for AgentTemplate, Harness, and ModelConfig resources using Kubernetes metadata. Pull request #2710
- Helm chart now supports global values for image registry, pull config, and namespace scope across all subcharts. Pull request #2817
Why it matters
V1.0.0-alpha1 marks a significant milestone with improved scalability for large clusters, better credential handling, and refined UI workflows. Security improvements include scoped authorization and proper egress policy provisioning, while Go and Python runtimes achieve feature parity on context compaction and Azure model support.
Changelog entry
- V1.0.0-alpha1 released Release
- Go runtime honors context compaction settings Pull request #2790
- AgentInstance lifecycle execution serialized to prevent duplicate mutations Pull request #2857
- Substrate gateway credential injection replaces literal environment variables Pull request #2868
- Substrate actors and workers support server-side pagination and search Pull request #2771
- Egress policies provisioned with Substrate v0.0.30 Pull request #2839
- Scoped authorization for AgentTemplate, Harness, and ModelConfig Pull request #2710
- Python runtime adds Azure foundry model support Pull request #2512
- Helm chart adds global values for registry and namespace scope Pull request #2817
- UI snapshots now named with Fork, Rename, Delete actions Pull request #2847
V1.0.0-alpha1 ships with server-side pagination for large inventories, gateway credential injection, scoped authorization, and runtime parity improvements across Go and Python.
Kagent v1.0.0-alpha1 is out. This release brings significant infrastructure improvements: server-side pagination and search for large cluster inventories, secure gateway credential injection replacing environment variables, scoped authorization for catalog resources, and full feature parity between Go and Python runtimes. The UI now supports named snapshots with inline actions, and Azure foundry models work across all compilers.