v0.0.8

AgentCore Launchpad v0.0.8 is the Console V2 baseline release. It is the last main release before the console-v3 line (Agent-DLC + V3 console, currently on console-v3) merges. It adds Agent IDE, a gateway mode for the COA context layer, read-back agent code and Runtime logs, and continuous evaluation inside Evaluation Tasks. It also fixes evaluation memory isolation and knowledge-base citations in chat.

Highlights

• Agent IDE (fifth build method): at /v2/coding, a member works with a codex coding assistant that writes, runs and tests code inside an AgentCore microVM, using per-user projects that persist across sessions. The project then deploys through the existing byoc pipeline as an ordinary governed agent. The IDE puts the conversation in the center, with collapsible and resizable Files and Workbench panels, plus file and project downloads.
• Context layer gateway mode: an admin can enable one platform-managed AgentCore Gateway per workspace for COA. Harness and Strands agents can then choose via Gateway instead of the default direct mode, and still act as the signed-in member, with the same consent card, preflight and grants. The agent wizard's Gateway picker shows the managed gateway. Members without a COA account can try the context layer with a shared, read-only COA demo account. A Strands turn now stops at the consent card while consent is pending.
• Agent code view: agent detail gains a Code page that reads back what AWS actually runs. For Runtime agents that is the S3 package, build context or staged BYOC upload; for Harness agents it is the GetHarness configuration with sensitive fields filtered out and masked.
• Runtime logs: the V2 agent detail shows a Runtime's stdout/stderr (FilterLogEvents), with time, session and keyword filters. The log group is derived on the server, and other members' private sessions are never shown.
• Continuous evaluation in Evaluation Tasks: V2 Online Evaluation is folded into Evaluation Tasks as a Continuous tab, leaving one list, one detail page and one editor. Old links still work.
• Evaluation isolation: every dataset replay and persona scenario now uses its own memory actor (<agent>__eval-<run>-<n>), so memory extracted in one run no longer leaks into later runs' scores. The Memory console hides these evaluation actors and labels the partitions of deleted agents.
• Region-aware built-in evaluators: only the built-in evaluators the region actually lists are offered and accepted. A run naming a missing one is refused before any invoke.
• BYOC code_zip: an entrypoint that declares LAUNCHPAD_ATTACHMENT_CONTRACT = "v1" receives native image and PDF attachments from Chat. Bundles that ship ADOT now appear in Observability.
• Chat rendering: GPT citation markers render as source chips, and bold markers next to CJK text render as bold. Knowledge-base sources an agent cites open through a short-lived presigned link instead of a private-bucket URL. Footnotes stay in the page and get unique ids per answer.
• Multi-account spokes: the spoke role can now turn on Transaction Search and run batch evaluations in a spoke region.
• macOS hosts: PDF attachments now work where the OS refuses RLIMIT_AS.

Upgrade notes

• Run make bootstrap (idempotent) to apply the launchpad-base stack change for the managed COA gateway (#318).
• Multi-account workspaces must update each spoke's launchpad-workspace-role stack from infra/spoke/launchpad-workspace-role.yaml (#328), then re-run the bootstrap observability stage for that workspace.
• Agent IDE is opt-in: an admin installs its per-workspace Runtime from the console; nothing installs on startup. Its tables are created automatically on startup, and existing tables are not migrated.
• Gateway mode is opt-in, and direct mode is unchanged. Gateway-mode consent cannot be revoked from the platform, because AWS keeps that token in the gateway's vault slot.
• Evaluation runs created before this release read their transcripts from the shared default actor, as before.
• Follow the production runbook for ledger backup, in-flight work checks, and deployment.

Validation

Release target: 77d36b8389fc27dffa99083a2989b6f38c9ba5bf from main.

The full make verify gate passed in an isolated checkout with locked dependencies:

• 7,285 backend tests, 59 Agent IDE runtime (Node) tests, and 49 infrastructure tests.
• 836 frontend tests across 81 files, ESLint, TypeScript checks, and the Vite production build.
• Local lifecycle lint/syntax checks, English/Chinese translation parity (8,655 keys), and Chinese punctuation checks.

GitHub's Python and JavaScript/TypeScript CodeQL checks also passed. The existing non-blocking React Hooks lint warning in EvaluationExperiment.tsx remains.

What's Changed

• Add a remote web coding agent: codex app-server on AgentCore, per-user projects, V2 web IDE by @xiehust in #310
• Let members try the context layer with a shared, read-only COA demo account by @xiehust in #311
• Coding IDE: conversation in the center, collapsible/resizable panels, file + project downloads by @xiehust in #312
• Rename the coding console to "Agent IDE" by @xiehust in #313
• End a Strands turn at the consent card while the member's COA consent is pending by @xiehust in #315
• Design demo-video covers in the V2 console's light style by @xiehust in #316
• Show each agent's code (or Harness configuration) read back from AWS by @xiehust in #317
• Context layer gateway mode (1/4): platform-managed COA gateway by @xiehust in #318
• Context layer gateway mode (2/4): consent preflight, auth card, completion, grants by @xiehust in #319
• Context layer gateway mode (3/4): Harness agents via the managed gateway by @xiehust in #320
• Context layer gateway mode (4/4): Strands agents via the managed gateway by @xiehust in #321
• Show the managed COA gateway in the agent wizard's Gateway picker by @xiehust in #322
• Merge V2 在线评估 into 评估任务 as a 持续评估 tab by @xiehust in #326
• Fix PDF attachments on macOS hosts (refused RLIMIT_AS) by @brilliantwf in #325
• BYOC code_zip: native Chat attachments and ADOT observability by @brilliantwf in #323
• Evaluation: offer and accept only the built-in evaluators the region has by @brilliantwf in #324
• Show an agent's Runtime stdout/stderr on the V2 agent detail, without leaking private sessions by @alex-sample-code in #314
• Let a spoke turn on Transaction Search and run batch evaluations by @xiehust in #328
• Give every evaluation scenario its own memory actor by @xiehust in #329
• Render GPT citation markers and CJK bold in chat markdown by @xiehust in #330
• Open knowledge-base sources an agent cites instead of linking to a private bucket by @xiehust in #334
• Keep chat footnotes in the page and give each answer its own footnote ids by @xiehust in #335
• Memory console: hide evaluation actors and tell deleted agents from their namesakes by @xiehust in #337

New Contributors

• @brilliantwf made their first contribution in #325

Full Changelog: 0.0.71...v0.0.8

0.0.71

AgentCore Launchpad 0.0.71 adds a governed COA context layer and improves Runtime canaries, trace message visibility, and dependency build reliability.

Highlights

• COA context layer: connect a Context Ontology Accelerator deployment, authorize access as the signed-in member, and attach governed context to Harness or Strands agents. The V2 Context Layer page includes namespace sources, metrics, ontology classes, deployment health, and a conversational playground.
• Runtime canaries: dataset replay sends each question to both arms, and verdicts require the minimum sample count in each arm. Runtime canaries can start at 50/50 and complete early from an eligible paired verdict. Custom evaluator scores also appear in per-case run comparisons.
• Trace messages: show recorded model inputs, completions, MCP arguments/results, and embedded Strands tool results on their own spans, with deduplication and session privacy boundaries preserved.
• Gateway user identity: logged-in Harness chat can forward the member's token to additional CUSTOM_JWT Gateways that meet the workspace trust checks, enabling tools that act as the user.
• Runtime rebuilds: converted Harness agents can recover from stale dependency pins when lock compilation fails. Rebuilding legacy Studio code removes obsolete generated mem0_memory imports and tool-list entries that otherwise prevent startup.
• Dependency validation: reject requirements-file option injection hidden after environment markers before resolving pins or compiling a lock, while retaining ordinary markers and supported URL/VCS pins.

Upgrade notes

• COA integration requires a compatible COA deployment, registration of each environment's Connection callback, and member consent. COA-attached agents currently require console user identity; the public API, evaluation replay, experiments, canaries, and Harness-to-Runtime conversion are unsupported for these agents.
• Additional Gateways must satisfy the workspace Cognito trust checks to receive a member token. Public API and evaluation invocations continue to use their service identity.
• Existing converted Runtime and legacy Studio agents receive the build fixes on their next redeploy or canary build.
• Follow the production runbook for ledger backup, in-flight work checks, and deployment.

Validation

Release target: 40f10c938175a885dc9c9fcc1647d575960a553b from main.

The full make verify gate passed in an isolated checkout with locked dependencies:

• 6,782 backend tests and 48 infrastructure tests.
• 653 frontend tests across 60 files, ESLint, TypeScript checks, and the Vite production build.
• Local lifecycle lint/syntax checks, English/Chinese translation parity (8,261 keys), and Chinese punctuation checks.

GitHub's Python and JavaScript/TypeScript CodeQL checks also passed. The existing non-blocking React Hooks lint warning in EvaluationExperiment.tsx remains.

What's Changed

• Forward the user token to other Gateways that trust the workspace pool by @xiehust in #296
• Show model and tool messages on trace spans by @xiehust in #297
• Let a Runtime canary complete with Dataset replay by @xiehust in #299
• Let agents use COA as a governed context layer, acting as the signed-in member by @xiehust in #302
• Browse and try COA context from the V2 Context Layer page by @xiehust in #303
• Drop mem0_memory from Studio code generated before 0b38065 (#301) by @xiehust in #304
• Re-resolve a converted Runtime's stale requirement pins when its lock fails (#300) by @xiehust in #305
• Reject requirements-file option injection in pinned dependencies by @xiehust in #308

Full Changelog: v0.0.7...0.0.71

v0.0.7

AgentCore Launchpad v0.0.7 adds per-agent evaluation history and comparison, per-person cost controls, and a more guided V2 console experience.

This release was updated to commit c49039dd0134ac18864c084990b61efbd4aee66c, adding #294 and #295 after the initial v0.0.7 publication.

Highlights

• Evaluation history and comparison: open Evaluation Tasks on per-agent scorecards, follow trends on the same benchmark, inspect version-by-evaluator summaries, and select two runs for overall or aligned case-by-case comparisons. Re-run eligible evaluations, filter the agent list, and preview session inputs and outputs in the trajectory picker. The architect's step 2 now shows run history, evaluator breakdowns, and changes by agent version using the same scorecard rules.
• Spend, budgets, and alerts: review spend by agent and person, configure threshold alerts with optional webhook delivery, assign monthly budget levels, and inspect personal usage. Long-running work records its payer and stops at supported safe boundaries when the budget is exhausted. Reconciliation replaces estimates with measured usage and accounts for AgentCore fees.
• Agent configuration history: inspect saved configuration snapshots, compare changes, and roll back through the deployment pipeline. The V2 create flow adds deployment progress, integration snippets, a quick chat, inline knowledge-base setup, and URL-backed filters.
• Model administration: manage deployment-wide model defaults and price overrides from the console, reset defaults, and probe model availability.
• Console navigation and feedback: use an essentials-only sidebar, command palette, glossary, help menu, launch checklist, and guided tour. Attention panels can be dismissed; chat ratings feed negative-feedback review in Insights. Runtime canaries are now the first A/B experiments tab.
• Reliability: observability counts AgentCore traces once, attention panels suppress stale evaluation failures, alert re-checks tolerate concurrent rule deletion, and transient-retry tests isolate their own evaluation runs.

Upgrade notes

• The Environments page and its compare/drift API have been removed. Agent configuration snapshots, diffs, and rollback remain available on agent details.
• Budget enforcement remains inactive until an administrator defines budget levels. Before assigning limits, configure prices for the models in use and review estimates for runtimes that do not report usage.
• Budget reconciliation is enabled by default and performs billable CloudWatch queries. Set LAUNCHPAD_BUDGETS_RECONCILE_ENABLED=false if reconciliation should remain disabled. Alert webhook scheduling is opt-in.
• Historical runs are assigned inferred snapshots only when deployment history supports the match. Runs older than all snapshots stay unknown when an intervening deployment succeeded, avoiding a false assignment to version #1. Per-case comparison depends on readable, alignable benchmark data; otherwise the console compares overall scores.
• Follow the production runbook for ledger backup, in-flight work checks, and deployment.

Validation

Release target: c49039dd0134ac18864c084990b61efbd4aee66c from main.

The full make verify gate passed in an isolated checkout with locked dependencies:

• 6,377 backend tests and 48 infrastructure tests.
• 603 frontend tests across 56 files, ESLint, TypeScript checks, and the Vite production build.
• Local lifecycle lint/syntax checks, English/Chinese translation parity, and Chinese punctuation checks.

The existing non-blocking React Hooks lint warning in EvaluationExperiment.tsx remains.

What's Changed

• V2 Observability ranks what needs a look and opens the failed traces by @xiehust in #268
• V2 list pages open with what needs attention: seven more modules by @xiehust in #269
• V2 shell: essentials-only sidebar, glossary, command palette and help menu by @xiehust in #270
• V2 agent create flow: deploy line, URL filters, integration snippets, try-chat, inline KB by @xiehust in #271
• V2 onboarding: launch checklist and a guided tour by @xiehust in #272
• Rate V2 Chat answers with a thumb; Insights lists the thumbs-down by @xiehust in #273
• Observability counts AgentCore traces, each once by @xiehust in #274
• Platform models become admin-overridable settings; prices cover the offered models by @xiehust in #275
• V2 spend by agent and person, and threshold alert rules with an opt-in webhook check by @xiehust in #276
• V2 agent config snapshots with diff and rollback, and environment compare and drift by @xiehust in #277
• Alert re-check survives a rule deleted mid-pass; offer GPT-6.1 Sol and GLM 5.3 by @xiehust in #278
• Hide the drift check from the V2 Environments page by @xiehust in #280
• Remove the Environments page and its compare/drift API by @xiehust in #281
• Per-person budgets (core): levels, usage ledger, prices, a hard-limit gate on every billable route, My usage by @xiehust in #282
• Per-person budgets for long-running jobs: payers, exact usage, estimates, safe-boundary stops by @xiehust in #284
• Per-person budgets: reconciliation and AgentCore fees (Runtime compute with idle tail, Memory, Gateway, Web Search, evaluations) by @xiehust in #285
• Eval transient-retry tests count only their own run (fixes a flaky verify gate) by @xiehust in #286
• Open the V2 architect's created datasets and evaluators on V2 pages by @xiehust in #283
• Show each session's input and output in the task wizard's Agent-trajectory picker by @xiehust in #287
• Evaluation scorecards backend: version stamps, scenario map, scorecard / compare / re-run routes by @xiehust in #289
• Compare two evaluation runs case by case in the V2 console by @xiehust in #290
• Open evaluation tasks on a per-agent scorecard in the V2 console by @xiehust in #291
• Dismiss V2 attention items, and stop listing stale evaluation failures by @xiehust in #292
• Put the runtime-canary tab first on the V2 A/B experiments page by @xiehust in #288
• Filter the 按 Agent table, hiding deleted agents by default by @xiehust in #293
• Show the architect's step 2 runs per version, on the scorecard rules by @xiehust in #294
• Leave a run's version unknown when the agent was redeployed before its first snapshot by @xiehust in #295

Full Changelog: v0.0.6...v0.0.7

v0.0.6

v0.0.6 releases commit 43e13c10cbe5d7a443c7481157eaf5e0be4e5e7c and includes the changes since v0.0.5.

Highlights

• Console V2 with native agent, evaluation, observability, governance, and administration pages, plus shared announcements and module tutorial videos.
• A solution architect assistant with reviewed deployment proposals, evaluation asset creation, prompt recommendations, and Harness version canaries.
• Skill Lab task generation, evaluation, training, multi-file inputs, artifact inspection, and publication to the Registry.
• Expanded evaluation workflows: online evaluation, custom Lambda evaluators, simulated personas, CloudWatch log datasets, prompt optimization, and paired canary scoring.
• BYOC agent creation, managed session storage, AgentCore Identity integration, inbound JWT, and outbound user authorization.
• Resource creator tracking and workspace isolation, consistent ledger backups, dependency security updates, and fixes for permission overrides and timestamp display.

What's Changed

• feat(evaluation): use native Claude SDK telemetry by @xiehust in #31
• feat(skill-lab): foundation — vendored SkillOpt, exec-worker infra, Bedrock judge by @xiehust in #32
• feat(skill-lab): task sets — entity, CLI-parity validation, console surface by @xiehust in #36
• feat(skill-lab): evaluation jobs — runner, results console, Registry deep link by @xiehust in #37
• feat(skill-lab): training jobs — gated optimization, diff review, publish to Registry by @xiehust in #38
• feat(skill-lab): Codex CLI target backend + judge model widening by @xiehust in #39
• feat(skill-lab): AI taskset generation (taskgen job type) by @xiehust in #40
• feat(skill-lab): unified select styling + built-in demo samples by @xiehust in #41
• feat(skill-lab): agentic judge (host-side, upstream parity) by @xiehust in #42
• feat(evaluation): skill evaluators + third-party/derived evaluators by @xiehust in #43
• fix(frontend): throw on unparseable 200 bodies instead of resolving null by @xiehust in #44
• fix(frontend): API cache poisoning — no-store on backend calls + body snippet in invalid-JSON errors by @xiehust in #45
• feat(frontend): DiffPanes renders a line-aligned split diff by @xiehust in #46
• feat(skill-lab): default judge gpt-5.6-sol + codex-backed agentic judge by @xiehust in #47
• feat(skill-lab): multi-doc training via trainable_files by @xiehust in #48
• fix(skill-lab): resolve uv alias symlinks in the venv interpreter by @xiehust in #49
• fix(workspaces): default-workspace resource merges mirror into the yaml by @xiehust in #50
• feat(skill-lab): binary and multi-file task assets by @xiehust in #51
• feat(skill-lab): accept md/txt/csv as task input files by @xiehust in #52
• feat(skill-lab): taskgen input attachments by @xiehust in #53
• fix(skill-lab): agentic judge routing, readiness, and diagnosability by @xiehust in #54
• fix(evaluation): preflight ground-truth judges and surface batch failure causes by @xiehust in #56
• feat(evaluation): per-agent online evaluation configs (?view=online) by @xiehust in #59
• feat(evaluation): online scores on session detail + Overview online-quality tile (P3a) by @xiehust in #60
• feat(evaluation): insights-mode online configs + recurring reports (P3b) by @xiehust in #61
• feat(optimization): pluggable recommend providers + gepa_lite reflection by @xiehust in #64
• feat(optimization): gepa_lite also optimizes the agent's own tool descriptions by @xiehust in #65
• fix(optimization): truncation-aware reflection retry; regenerate keeps the pinned run by @xiehust in #66
• fix(optimization): accept defaults to the recommended tool descriptions; tolerant reflection JSON by @xiehust in #68
• chore(optimization): default reflection model → Claude Opus 5 by @xiehust in #69
• fix(experiment): remember the recommend pickers per experiment by @xiehust in #70
• fix(optimization): stream the reflection call with a long read timeout by @xiehust in #71
• Catch-all route: unknown URLs render a not-found view inside the shell by @xiehust in #72
• Unreachable backend: health chip reflects reality and list pages show load errors, not empty states by @xiehust in #73
• Stale "Phase" copy and unlabeled form controls by @xiehust in #74
• Small viewports: no horizontal page scroll below 720 px by @xiehust in #75
• AWS ClientErrors on console routes map to 4xx envelopes, never a bare 500 or raw boto text by @xiehust in #76
• Disabled primary actions explain what is missing (shared Btn disabledReason) by @xiehust in #77
• Stale deep links tell the user the resource is gone instead of silently falling back by @xiehust in #78
• zh-CN typography: full-width punctuation inside Chinese copy, consistently by @xiehust in #79
• Evaluation runs can be stopped (StopBatchEvaluation + queued cancel) by @xiehust in #80
• Agent detail lists AWS versions and endpoints (Runtime + Harness) by @xiehust in #81
• Chat can end the AgentCore Runtime session (StopRuntimeSession) by @xiehust in #82
• Memory resources: edit description and event expiry (UpdateMemory) by @xiehust in #83
• Governance gateway detail manages Gateway rate limits by @xiehust in #84
• Cloud datasets: re-sync edits the AWS draft in place; drafts show status, publish and list versions by @xiehust in #85
• Cloud dataset runs can pin a published dataset version by @xiehust in #86
• Governance gateway detail can synchronize dynamic MCP-server targets (SynchronizeGatewayTargets) by @xiehust in #87
• A2A registry records show the live agent card served by the runtime (GetAgentCard) by @xiehust in #88
• Bootstrap memory carries session-summary and episodic (reflection) strategies by @xiehust in #89
• Chat memory rail: list_events accepts the pinned memory_id again by @xiehust in #90
• Governance gateway detail names each target's kind (MCP / HTTP / inference variants) by @xiehust in #91
• Evaluators: create, show and safely edit code-based (Lambda) custom evaluators by @xiehust in #92
• Registry consumer view: what ListDiscoverableRegistryRecords exposes vs the control-plane list by @xiehust in #93
• A2A runtime card and registry record agree on version by @xiehust in #94
• Observability session detail can score a session on demand (data-plane Evaluate) by @xiehust in #95
• Default model and Registry-GA facts corrected in troubleshooting, setup and README (en + zh) by @xiehust in #96
• AGENTS.md re-mirrors CLAUDE.md; both fix the route list and the make dev port claim by @xiehust in ht...

v0.0.5

What's Changed

• refactor(aws): funnel all boto3 construction through a workspace-keyed factory by @xiehust in #25
• feat(workspaces): multi-workspace core — per-environment boundary, RBAC, and console bootstrap (P2) by @xiehust in #26
• feat(workspaces): cross-account workspaces via AssumeRole + spoke role template (P3) by @xiehust in #27
• docs(workspaces): feature summary in the READMEs, operational notes in both runbooks by @xiehust in #28
• feat(workspaces): purge, registration preflight, in-page setup guide, grants table by @xiehust in #29
• chore: remove unrelated local workshop content committed by mistake by @xiehust in #30

Full Changelog: v0.0.41...v0.0.5

v0.0.41

Full Changelog: v0.0.4...v0.0.41

v0.0.4

What's Changed

• feat(eval): run up to 3 batch evaluations concurrently — the eval queue previously serialized all runs on a one-active-batch assumption; the actual AWS quota is 5 active batch evaluations per account. Runs now execute through a bounded-concurrency queue (eval_max_concurrent_runs, default 3, clamped 1–5), excess runs still queue instead of failing, and StartBatchEvaluation retries transient Throttling/Conflict/ServiceQuotaExceeded errors before failing a run.
• feat(auth): starting evaluation/insights runs is now member-grantable (perm:eval.run) — granted to every member by default and revocable per user in the Users console, instead of flat administrator-only.
• chore(deps): resolved all open Dependabot alerts — dompurify → 3.4.13 and prismjs → 1.30.0 via npm overrides, js-yaml → 4.3.1, h2 → 4.4.1, plus an npm audit fix sweep (brace-expansion, fast-uri, find-my-way, nanoid, postcss).
• fix(kb): return KB creation immediately instead of polling toward ACTIVE (#20, fixes #19)
• fix(local-exec): apply RLIMIT_NPROC only under the dedicated exec user
• fix(studio): drop mem0 and sampling params from generated code; UI polish
• fix(ui): wrap long lines in the deploy job log panel

Full Changelog: v0.0.3...v0.0.4

v0.0.3

What's Changed

• Security: production boundary, supply-chain pinning, and per-agent execution roles by @xiehust in xiehust/agentcore_launchpad#9
• fix(deploy): retry IAM propagation on the zip and harness paths too by @xiehust in xiehust/agentcore_launchpad#10
• feat(evaluation): connect trace readiness to Bundle A/B experiments by @xiehust in xiehust/agentcore_launchpad#11
• fix(canary): stop rendering a partial setup artifact as a complete one by @xiehust in xiehust/agentcore_launchpad#12
• fix(convert): make harness→zip conversion actually deployable and invocable by @xiehust in xiehust/agentcore_launchpad#13
• docs(spec): record the conversion contracts PR #13 established by @xiehust in xiehust/agentcore_launchpad#14
• feat(runtime): attach registry skills to zip agents by @xiehust in xiehust/agentcore_launchpad#15
• fix(evaluation): orient A/B verdict by evaluator polarity, reject ground-truth judges online by @xiehust in xiehust/agentcore_launchpad#16
• perf(experiments): parallelize gateway traffic send with bounded concurrency by @xiehust in xiehust/agentcore_launchpad#19

Full Changelog: xiehust/agentcore_launchpad@v0.0.1...v0.0.3