AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
Agent Advisor provides a structured, multi-phase workflow to help users determine the optimal AWS runtime for their AI agents. It uses a deterministic scoring engine to evaluate requirements and guides users through a process that includes discovery, architectural design, cost estimation, and the creation of an executable proof-of-concept. The skill supports various AWS compute options, including AgentCore, ECS, EKS, and Lambda, and integrates with migration engines for existing workloads.
What it does
- Recommends AWS runtimes for AI agents based on technical requirements.
- Generates migration plans for existing AI workloads moving to AWS.
- Provides architectural guidance for agentic systems.
- Assembles executable proof-of-concept scaffolding.
- Estimates coarse cost magnitudes for proposed agent deployments.
- Integrates with external migration engines for workload transitions.
When to use it
- When you need to decide which AWS runtime is best for a new AI agent.
- When migrating an existing AI agent workload to AWS.
- When you need to add services like memory, identity, or observability to an existing AWS agent.
- When you require a structured plan and POC for deploying agents on AWS.
Advantages
- Provides a deterministic, repeatable recommendation process.
- Supports a wide range of AWS compute environments.
- Offers a unified flow from initial idea to executable POC.
- Adapts to both technical and business-focused user backgrounds.
Usage examples
- “I have an agent idea, what do I build?”
- The agent initiates the intake process to clarify your requirements and recommend an appropriate AWS architecture.
- “Move my agents to AWS.”
- The agent triggers the migration workflow to assess your existing workload and plan the transition to an AWS runtime.
- “Which runtime for my agent, AgentCore vs ECS vs EKS vs Lambda?”
- The agent runs a scoring analysis to compare these runtimes against your specific agentic needs.
- “Add AgentCore services to an agent already on AWS.”
- The agent enters the add-capabilities branch to guide you through integrating memory, gateway, and observability services.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/aws/agent-advisor/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip agent-advisor.zip -d .agents/skills
Documentation by AWS
AWS Agent Advisor
Helps startups decide how and where to run AI agents on AWS. Deterministic scoring recommends a runtime; the conversation adapts to the user's technical background.
Definitions
"Load" = Read the file with the Read tool and follow it. Do not summarize or skip.
$RUN_DIR= the run directory under.agent-advisor/(e.g..agent-advisor/0630-1430/), created in Intake.$PLUGIN=${CLAUDE_PLUGIN_ROOT}(the installed plugin root). On Claude Code this token substitutes inline. If${CLAUDE_PLUGIN_ROOT}does not resolve (some Cursor/Codex builds, or a literal${CLAUDE_PLUGIN_ROOT}string showing up in a path error), fall back to the skill's own directory: this SKILL.md lives at<plugin>/skills/agent-advisor/SKILL.md, so the engine and its data are all inside this skill — scripts at./scripts/..., runtime profiles at./references/runtimes/..., and decision refs at./references/decision-refs/...relative to it. Prefer${CLAUDE_PLUGIN_ROOT}/skills/agent-advisor/...; use the relative fallback only when it fails to resolve.
Prerequisites
uvavailable (for scoring). Check:uv --version. If missing, tell the user to install it from the official install guide (https://docs.astral.sh/uv/getting-started/installation/ — e.g.brew install uvorpipx install uv) and stop.
Phase Structure (frontmatter)
Phase, fragment, and assembler files carry a YAML frontmatter block that declares how each
phase is composed — its inputs, triggers, fragments, assembler, artifacts, gates, and
ordering. The execution contract is the vendored references/vendored/dsl/INTERPRETER.md:
it defines every frontmatter key, the fragment/assembler model, the gate protocol
(HANDOFF_OK / GATE_FAIL), and the interpreter loop. Load it first (once, at the
start of a run), then execute each phase file's prose body. Elsewhere in this skill,
INTERPRETER.md (without a path) refers to this loaded contract.
Execution
This skill is driven by the interpreter loop in INTERPRETER.md (§ The interpreter loop):
it reads .phase-status.json, determines the current phase, runs each phase's
_preconditions / fragments / _assemble / _postconditions, advances on HANDOFF_OK
via _advances_to, and validates state. The backbone (intake → discover → clarify →
confirm → design → estimate → generate → migration-plan → poc → complete) and the
one sidebar branch (add-capabilities) are derived
from the phase files' frontmatter — they are not restated here.
Cold start (entry phase). With no run under .agent-advisor/ carrying a
.phase-status.json, begin at references/phases/intake/intake.md — this skill's entry
phase (the one carrying _init: true). On a warm start, current_phase in
.phase-status.json is authoritative (INTERPRETER.md § The interpreter loop).
Skill bindings (INTERPRETER.md § Skill bindings). This skill declares:
Run root:
.agent-advisor/—$RUN_DIRis this skill's name for the run directory (.agent-advisor/[MMDD-HHMM]/). Intake's own prose performs the_initbootstrap.State shape: § State file below (advisor-specific keys such as
entry_point,audience,recommendation_reviewed,migration_plan_ctx,migration_plan_unavailable); the shared state schema is not vendored.Run seed (optional):
$RUN_DIR/seed.json, else.agent-advisor/seed.jsonat the run root (schemascripts/schemas/seed.json) supplies machine-readable answers for a non-interactive run — the Clarify dimensions, the two gate answers, the POC mode, the live-probe answer, and aco_recommendtie-break. It is the HIGHEST-precedence source for every value it carries (clarify.md Step 2.5), which is what makes a repeated run's score comparable: the deterministic engine gets byte-identical input. A gate the seed omits is declined; a dimension the seed omits falls through to detection, then prose, then anassumedvalue that MUST be recorded in$RUN_DIR/UNANSWERED.md. With no seed, the interactive flow is unchanged.Resolved statuses:
skipped(routing resolved the phase without running it), plusnot_applicableformigration_planonly.Conditional backbone routing: the entry-point routing below. When a routing rule marks a phase not-applicable, set it
skippedand advance through its_advances_toin the same state write.
Routing & gates (orchestration)
Sidebar placement and conditional backbone routing are orchestration prose owned by
this file (INTERPRETER.md § Skill bindings, § Backbone vs sidebar).
Entry-point routing:
build_scratch→ skip Discover; Clarify → Confirm → Design → Estimate → Generate → Gate 2 → POC (any winning runtime). No migration plan (nothing existing to migrate).build_deploy→ Discover (if code) → Clarify → Confirm → Design → Estimate → Generate → Gate 1 → Migration Plan (if existing non-AWS AI workload detected and user confirms) → Gate 2 → POC (any winning runtime).migrate→ Discover (if code) → Clarify → Confirm → Design → Estimate (target-state run cost; migration TCO comparison stays with the Migration Plan engine) → Generate → Gate 1 → Migration Plan (in-skill, reusing the siblinggcp-to-awsskill) → Gate 2 → POC (any winning runtime, when the plan was produced). Declining Gate 1 keeps the classic handoff: pointer to/aws-startup-advisor:llm-to-bedrockwithhandoff-summary.md.add_capabilities→ loadreferences/phases/add-capabilities/add-capabilities.mdand follow it (no runtime scoring; writescapabilities-recommendation.md). This is a self-contained branch — it does NOT pass through Clarify / Confirm / Design / Estimate / Generate, so the phase gate below never applies to it.Temporal detection routes into
migratewith temporal units pre-seeded (see discover).
Gate semantics (backbone tail):
Gate 1 →
migration_planruns only whengenerateis done ANDrecommendation_reviewed == true(generate.md Step 5.5) AND entry point ∈ {migrate, build_deploy} AND the run is migration-eligible (generate.md Step 6) AND the user confirmed Gate 1. Otherwise resolve it:not_applicable(build_scratch / no migratable workload) orskipped(declined) — and advance.Gate 2 →
pocruns only whenphases.poc == "in_progress"(set when the user answers Gate 2 "yes" — asked in generate.md Step 7 or migration-plan.md Step 6) ANDrecommendation_reviewed == true. Any winning runtime (agentcore / ecs / eks / lambda / lambda_microvms) — the POC shape follows the verdict (poc.md Step 3 dispatch onreferences/decision-refs/poc-shapes.md). Gate 2 is only offered whenmigration_plan∈ {completed, skipped, not_applicable} — orin_progresson build_deploy only (Stage 2 failed/aborted; fallback POC from design.json per migration-plan.md failure handling); for entry pointmigrate, only whenmigration_plan == "completed"(the POC implements the plan) OR when the stage resolvednot_applicablewithmigration_plan_unavailable == "engine_absent"— a standalone deployment that does not bundle the migration engine, where Gate 2 is offered by migration-plan.md Step -1 and the POC is design-backed. A migrate-POC with no plan for any OTHER reason (the user declined) has nothing to implement.Persisting Gate 2 as
phases.poc = "in_progress"BEFORE poc.md loads makes the confirmation resumable: if the session breaks between the "yes" and the load, the interpreter re-enterspocwithout re-asking. (A declared deviation fromINTERPRETER.md§ The interpreter loop step 5's gate-then-in_progressordering — the user's confirmation is the entry event worth persisting.)
Phase gate: Do NOT load design.md / estimate.md / generate.md unless
$RUN_DIR/.phase-status.json exists and BOTH phases.clarify == "completed" AND
phases.confirm == "completed". Confirm confirms the deployment model, the service
set, and (for a co_recommend tie) the user's chosen_runtime — Design and the diagram depend on
its confirm.json output, so it must not be skipped. If the user asks to skip Clarify or Pass 2,
refuse briefly and run it.
State file (.phase-status.json)
{
"run_id": "0630-1430",
"entry_point": "build_scratch",
"audience": "technical",
"current_phase": "clarify",
"phases": {
"intake": "completed",
"discover": "skipped",
"clarify": "in_progress",
"confirm": "pending",
"design": "pending",
"estimate": "pending",
"generate": "pending",
"migration_plan": "pending",
"poc": "pending"
}
}Status values: pending → in_progress → completed, plus skipped. Use read-merge-write:
read before each update, change only the advancing keys, keep prior phases.
recommendation_reviewed (top level, boolean) is set to true by generate.md Step 5.5 when
the user explicitly confirms they have seen the recommendation. Gate 1, Gate 2, and the
migration_plan / poc states all require it — no gate may be asked while it is absent.
migration_plan additionally uses not_applicable (build_scratch, or no migratable workload
detected). When Stage 2 runs, migration_plan_ctx is added at the top level:
{"repo": "<abs path to target repo>", "migration_dir": "<abs path to .migration/<id>/>"} —
Stage 3 reads gcp-to-aws artifacts ONLY via this recorded path, never by re-globbing.
Files
| File | Purpose |
|---|---|
references/vendored/dsl/INTERPRETER.md | Vendored DSL execution contract (interpreter loop + gate protocol) |
references/phases/intake/intake.md | Entry point + technical background + open context |
references/phases/discover/discover.md | Lightweight code detection |
references/phases/clarify/clarify.md | Clarify orchestrator + answer mapping to scoring keys |
references/phases/clarify/clarify-technical.md | Technical-background question wording |
references/phases/clarify/clarify-business.md | Business-background question wording |
references/phases/confirm/confirm.md | Winner-specific follow-ups |
references/phases/design/design.md | Assemble recommendation; Migrate handoff branch |
references/phases/estimate/estimate.md | Coarse cost magnitude |
references/phases/generate/generate.md | Layered recommendation doc + scaffolding |
references/phases/migration-plan/migration-plan.md | Stage 2: full migration plan via the sibling gcp-to-aws engine |
references/decision-refs/temporal.md | Temporal rules: Tier 1/2 tables, adapter, runbooks, commercials (consumed by discover/clarify/design/generate) |
references/decision-refs/poc-shapes.md | Per-runtime POC deploy shapes (ECS/EKS/Lambda/MicroVMs/Temporal) |
references/decision-refs/*.md | Runtime service cards, model defaults, freshness |
references/decision-refs/workload-classes.md | Deterministic verdicts for non-agent workload units (batch/service/io) |
references/runtimes/*.json | Runtime registry (read by scoring.py) |
scripts/scoring.py | Deterministic scoring engine |
scripts/test_temporal_decision_refs.py | Content lock for the Temporal decision reference |
scripts/test_poc_shapes.py | Content lock for the POC deploy shapes |
scripts/test_workload_classes.py | Content lock for workload-classes.md (verdicts table) |
scripts/test_unit_grouping.py | Unit grouping + pattern matching (workload-class assignment) |
scripts/test_collapse_invariant.py | Collapse-invariant ordering enforcement (A→B implies [B] ⊆ [A] outputs) |
Maturity and readiness contract
Intake persists target_maturity (prototype, private_beta, or production) in run state. Clarify carries it and the readable readiness record into answers.json. Current-run verification evidence remains exclusively in the sibling $RUN_DIR/current-run-verifications.json artifact: never copy it into seed.json or answers.json. Design and Generate may consume that artifact only after validating its schema and matching run_id, and may carry forward verified outcomes but not the raw evidence records. Load references/decision-refs/maturity-readiness.md whenever target maturity is selected. A cached volatile fact may inform discovery, but only a record verified in this run can make a verification-required constraint final; otherwise the score remains provisional with deferred verification requirements.
Contextual offers (final step)
Before you finish: your answer is not complete until you have run the Contextual offers (final step) below.
This is a mandatory closing step of every final recommendation — complete it before you consider your answer done. Run this offer check inline yourself; do not depend on loading another skill. The recommendation is decided on merit first; this step never changes, re-ranks, or re-opens it, and the offer is looked up only after the advice is final.
Read
../knowledge-base-for-startups/references/offers.md— the offer index (Offer · Summary · Keywords), one row per offer.Judge relevance against what the founder actually surfaced this session. Surface an offer only if one holds: a named-product match (you recommended a specific partner product by name and it has a row); a direct intent/component match (an offer's Keywords directly serve a concrete workload, a component in the final recommendation, or a need the founder raised). Keep a high bar — if the match is weak, or you would be inferring an unstated need, append nothing. Never surface an offer that competes with a product or provider the founder already uses or decided to keep (e.g. existing auth provider → no Auth0 unless they are already on Auth0, and then only as a credit on it; existing vector DB such as Pinecone → no Qdrant).
If (and only if) one offer clears the bar, open only its detail file
../knowledge-base-for-startups/references/offers/<slug>.mdand append exactly one quiet, dismissible italic line — stating the value and condensing eligibility faithfully from that file (drop no qualifying criterion; never fabricate), with the redeem link built from the canonicalsource_urlplus asourceattribution param — append?source=ide-startupAdvisor-<host>(or&source=...if the url already has a query string), filling<host>from your execution environment (Claude Code →claude; AWS Startup Advisor IDE extension / VS Code →vscode; Cursor →cursor; Kiro →kiro; Codex →codex), or the genericide-startupAdvisorwhen you cannot tell. Only append the query param; never change the destination. When an offer clears the bar you MUST append that single offer line as the last step; do not omit it. The offer is never the pick — only a line under the AWS pick — and stage gates in the references apply (e.g. no Datadog line for an early-stage founder; see../architect-for-startups/references/observability.md). For a single-service pick (auth, payments, vector DB, observability, etc.) where missing budget/team/traffic inputs would not change the pick, assume an early-stage (pre-PMF) startup unless the founder says otherwise and answer in one turn without a stage-only question; this only removes a stage-only question and never skips the Intake/Clarify/Confirm gates before final advice or the POC. Show exactly one, never a list. If none clears the bar, append nothing and do not mention offers, partner alternatives, or why none appeared.
Standard line:
***<Offer>** has an exclusive offer via AWS Activate — <value>, <eligibility>. [Redeem →](<source_url>?source=ide-startupAdvisor-<host>)*Caps and control: at most one offer per response and often none; no more than one per five messages and two per session; show a given offer at most once per session and never one already shown, claimed, or dismissed; if the founder has muted offers, skip this step entirely. These per-five-messages, per-session, and already-shown caps are session-state limits; in a fresh session with no prior offers they are non-binding, so do not withhold an otherwise-qualifying offer merely because you cannot verify session history. See ../contextual-offers-for-startups/SKILL.md for the full rules — but perform the check inline; it must not depend on that skill being loaded.
Files
- ATTRIBUTION.mdAdded by AutoGineer907 B
- LICENSELicence file added by AutoGineer9.9 KiB
- NOTICELicence file added by AutoGineer67 B
- SKILL.md19.2 KiB
- references/decision-refs/agentcore.md5.4 KiB
- references/decision-refs/batch.md2.3 KiB
- references/decision-refs/cost-levers.md1.8 KiB
- references/decision-refs/ecs.md1.4 KiB
- references/decision-refs/eks.md1.1 KiB
- references/decision-refs/freshness.md10.7 KiB
- references/decision-refs/lambda-microvms.md1.9 KiB
- references/decision-refs/lambda.md1.1 KiB
- references/decision-refs/managed-alternatives.md1.9 KiB
- references/decision-refs/maturity-readiness.md2.8 KiB
- references/decision-refs/model-selection.md4.2 KiB
- references/decision-refs/poc-shapes.md15.5 KiB
- references/decision-refs/temporal.md17.5 KiB
- references/decision-refs/workload-classes.md2.2 KiB
- references/diagram/build-diagram.md5.1 KiB
- references/handoff/handoff-migration.md2.1 KiB
- references/output-templates/recommendation-doc.md5.1 KiB
- references/phases/add-capabilities/add-capabilities-assemble.md951 B
- references/phases/add-capabilities/add-capabilities.md4.6 KiB
- references/phases/clarify/clarify-assemble.md1.4 KiB
- references/phases/clarify/clarify-business.md3.3 KiB
- references/phases/clarify/clarify-technical.md2.7 KiB
- references/phases/clarify/clarify.md26.5 KiB
- references/phases/confirm/confirm-assemble.md1019 B
- references/phases/confirm/confirm.md15.1 KiB
- references/phases/design/design-assemble.md1.7 KiB
- references/phases/design/design.md26.9 KiB
- references/phases/discover/discover-assemble.md1.2 KiB
- references/phases/discover/discover.md10.0 KiB
- references/phases/estimate/estimate-assemble.md780 B
- references/phases/estimate/estimate.md19.0 KiB
- references/phases/generate/generate-assemble.md1.0 KiB
- references/phases/generate/generate-report.md45.6 KiB
- references/phases/generate/generate.md15.7 KiB
- references/phases/intake/intake-assemble.md991 B
- references/phases/intake/intake.md5.6 KiB
- references/phases/migration-plan/migration-plan-assemble.md1.2 KiB
- references/phases/migration-plan/migration-plan-gcp-constraints.md3.9 KiB
- references/phases/migration-plan/migration-plan.md32.1 KiB
- references/phases/model-recommend/model-recommend-assemble.md937 B
- references/phases/model-recommend/model-recommend.md17.2 KiB
- references/phases/poc/poc-assemble.md1.0 KiB
- references/phases/poc/poc-report.md12.6 KiB
- references/phases/poc/poc.md44.8 KiB
- references/report-help-banner.md6.0 KiB
- references/report-shell.md12.3 KiB
- references/runtimes/agentcore.json3.1 KiB
- references/runtimes/ecs.json1.5 KiB
- references/runtimes/eks.json1.4 KiB
- references/runtimes/lambda.json2.4 KiB
- references/runtimes/lambda_microvms.json3.3 KiB
- references/vendored/README.md2.0 KiB
- references/vendored/ai/bedrock-pricing-cache.md57.2 KiB
- references/vendored/ai/models/anthropic-bedrock.json8.6 KiB
- references/vendored/ai/models/openai-bedrock.json28.0 KiB
- references/vendored/dsl/INTERPRETER.md51.8 KiB
- scripts/anthropic_model_recommendation.pyScript35.0 KiB
- scripts/build_diagram.pyScript21.2 KiB
- scripts/model_recommendation.pyScript8.8 KiB
- scripts/openai_model_recommendation.pyScript58.1 KiB
- scripts/pyproject.toml170 B
- scripts/schemas/current-run-verifications.json1.1 KiB
- scripts/schemas/model-recommendation-input.json5.4 KiB
- scripts/schemas/model-recommendation.json13.1 KiB
- scripts/schemas/model-verification.json2.5 KiB
- scripts/schemas/scoring-result.json3.5 KiB
- scripts/schemas/seed.json5.6 KiB
- scripts/score_units.pyScript2.9 KiB
- scripts/scoring.pyScript16.5 KiB
- scripts/test_build_diagram.pyScript21.8 KiB
- scripts/test_collapse_invariant.pyScript2.0 KiB
- scripts/test_cost_levers.pyScript1.7 KiB
- scripts/test_model_recommendation.pyScript13.9 KiB
- scripts/test_openai_model_recommendation.pyScript46.5 KiB
- scripts/test_opus55_generation_contracts.pyScript20.8 KiB
- scripts/test_opus55_integration.pyScript21.5 KiB
- scripts/test_poc_shapes.pyScript14.2 KiB
- scripts/test_scoring.pyScript35.3 KiB
- scripts/test_temporal_decision_refs.pyScript7.4 KiB
- scripts/test_unit_grouping.pyScript71.7 KiB
- scripts/test_verify_model_path.pyScript25.8 KiB
- scripts/test_workload_classes.pyScript5.2 KiB
- scripts/uv.lock97.8 KiB
- scripts/verify_model_path.pyScript10.9 KiB
references/phases/discover/discover.md
_phase: discover _title: "Discover — Lightweight Detection" _requires_phase: intake _input: workspace _assemble: _file: phases/discover/discover-assemble.md _produces:
context-signals.json _advances_to: clarify _preconditions:
_check_phase_completed: intake _on_failure: _halt_and_inform _postconditions:
_check_file_exists: context-signals.json _on_failure: _halt_and_inform
_validate_json: context-signals.json _on_failure: _halt_and_inform
_assert: "context-signals.json contains only signals detected with reasonable confidence, with a _detected array; model_provider recorded when detection succeeded; units[] is non-empty with legal workload_class, coupling.mode, and trigger values, kebab-case ids, and no in_process coupling (merged during grouping); when Temporal is detected, units include the temporal_worker_poll tier and per-Activity-class entries, and the temporal context object records server state; temporal_worker_poll units list their queues" _on_failure: _halt_and_inform
Phase: Discover — Lightweight Detection
Only runs for build_deploy / migrate when the user provided a code path. Stays independent (does NOT require a separate migration plugin).
Step 1 — Scan for signals (read-only)
In the provided path, look for:
Framework (imports / requirements.txt / package.json):
strands,langgraph/langchain,crewai/autogen,openai(Agents SDK), elsecustom/none.Model call inventory: for every model-bearing unit, record exact provider, raw model-ID strings, SDK/package, API surface, and source paths. Provider vocabulary:
anthropic | openai | azure_openai | google_genai | bedrock | none | unknown. API surfaces include Anthropicmessages, OpenAIchat_completions/responses, and Bedrockconverse/invoke_model. Do not collapse exact IDs such asclaude-3-7-sonnet-latestinto a coarseclaudefamily.Gateway/router detection (OpenRouter, LiteLLM): a
base_url(orOPENAI_BASE_URLenv var) containingopenrouter.ai, or anOPENROUTER_API_KEY/OR_API_KEYpresent, means the calls are routed through OpenRouter, not the provider's direct API. This is a transport difference, not a different model — classifyproviderfrom the underlying model, not asunknown. OpenRouter model IDs are provider-prefixed slugs (openai/gpt-5.4,anthropic/claude-sonnet-4-6,google/gemini-2.5-pro); the prefix names the realprovider. Record the raw slug unchanged inmodel_ids[](do not strip the prefix — Model Recommend's family detection tolerates it) and setsdkto"openai (via OpenRouter)"so the transport is visible downstream. The same rule applies to a detected LiteLLM router (litellmimport,LITELLM_PROXY_BASE_URL), withsdk: "openai (via LiteLLM)". Without this rule, a real OpenAI model routed through a gateway falls through toprovider: "unknown"and is scored as if it were an unclassified/Bedrock-native source.
Anthropic migration-sensitive features:
budget_tokens;temperature/top_p/top_k; assistant-prefill messages; refusal handling; explicit or framework-defaultmax_tokens; structured output; prompt caching; server tools; Files API / URL sources; Message Batches; fallbacks; conversation-state IDs; Skills/MCP connector/managed-agent dependencies. Record observed feature codes and source paths for Model Recommend.Session/timeout hints: timeout configs, long-running loops, queue/HITL patterns.
Multi-tenant hints: per-user/tenant scoping, separate contexts.
Compute hints: GPU instance types, heavy compute (compilation, ML inference).
Data store hints: Redis/DynamoDB/vector store connections.
Temporal orchestration:
temporalio(Python) /go.temporal.io(Go) /temporal-sdk(Java) /@temporalio/*(TypeScript) imports or dependencies.Unit inventory: enumerate the system's workload units. Signals:
agent graphs/swarms (LangGraph graph, CrewAI crew, custom loop) →
agent_sessioncron/scheduler entries, queue consumers doing periodic bulk work →
batchwebhook handlers, thin API routes, event handlers →
light_iolong-running servers / WebSocket daemons that are not agents →
service
Grouping rule (first, before writing units): agents that interact IN-PROCESS — nodes of one framework graph, members of one swarm, direct in-process calls — are ONE unit; they deploy and scale together, and that unit's
multi_agentscoring answer isyes. Workloads that interact across processes (queue,api,a2a) or not at all (none) are separate units.in_processtherefore never appears in a written unit'scoupling.mode— it is the merge criterion, not a link type. Closed vocabularies:workload_class∈ {agent_session,batch,light_io,service,temporal_worker_poll};coupling.mode∈ {queue,api,a2a,none}.Temporal detection (when the signals above are found): create one
temporal_worker_pollunit per worker fleet (task-queue group), plus one unit per Activity execution class. Activity classification uses the vocabulary inreferences/decision-refs/temporal.md(Tier 2): agent-session Activities →agent_session; batch Activities →batch; light-IO Activities →light_io. Worker registration patterns (Python:Worker.run()orclient.execute_workflow(), Go:worker.RegisterWorkflow, Java:@WorkflowMethod/@ActivityMethod, TypeScript:Worker.create({workflowsPath})) and task-queue names in config files or Worker constructors identify worker fleets. Eachtemporal_worker_pollunit gains"queues": ["<task queue names this unit polls>"](from the detected task_queues, grouped per unit). Each Activity-class unit (agent_session / batch / light_io created from an Activity) gains"task_queue": "<the task queue this Activity runs on>"— the queue that ties it to the worker fleet that executes it (from the Activity'stask_queue=inexecute_activity/proxyActivities/ Worker registration). This is the join key between a poll unit'squeues[]and its Activity units: a diagram or plan connects ONLY the fleet whosequeues[]contains an Activity'stask_queue, never a cartesian product across all fleets. With a single fleet on a single queue, every Activity carries that one queue. On Temporal detection, record the following detection signals:temporalio,go.temporal.io,@temporalio/*, ortemporal-sdkpackage imports; Worker registration calls (Worker.run(),worker.RegisterWorkflow,Worker.create()); task-queue names in source or config; Workflow decorators / annotations (@workflow.defn,@WorkflowMethod); Activity decorators / annotations (@activity.defn,@ActivityMethod);workflow.ExecuteActivitycall sites;proxyActivities<T>()(TypeScript); server connection config (*.tmprl.cloudor self-hosted address). Write a top-leveltemporalcontext object incontext-signals.json:{ "detected": true, "server": "cloud|self_hosted|unknown", "sdks": [], "task_queues": [] }.Trigger detection: each unit gains a
triggerfield capturing how it is invoked. Closed vocabulary:trigger∈ {request,event,schedule,temporal,unknown}. Detect from handler type (REST/WebSocket handlers →request), cron entries or scheduler configs →schedule, event-source wiring (queue consumers, S3 listeners, SNS/webhook handlers) →event, Temporal Activity registration →temporal.
Step 2 — Map to pre-filled answers
Write $RUN_DIR/context-signals.json mapping detected signals onto scoring keys, e.g.:
{
"framework": "langgraph",
"multi_agent": "yes",
"session_state": "hitl",
"model_provider": "openai",
"_detected": [
"framework from imports",
"multi_agent from graph with 2+ nodes",
"model_provider from SDK imports"
],
"units": [
{
"id": "chat-agent",
"workload_class": "agent_session",
"trigger": "request",
"source": "detected",
"framework": "langgraph",
"model_source": {
"provider": "anthropic",
"model_ids": ["claude-3-7-sonnet-latest"],
"sdk": "anthropic",
"api_surface": "messages",
"source_paths": ["src/agent.py"],
"detected_features": ["budget_tokens", "assistant_prefill"]
},
"coupling": {
"interacts_with": ["summarizer"],
"mode": "queue"
},
"evidence": "LangGraph StateGraph with 3 nodes; pushes summaries to Redis queue"
},
{
"id": "summarizer",
"workload_class": "batch",
"trigger": "schedule",
"source": "detected",
"coupling": {
"interacts_with": [],
"mode": "none"
},
"evidence": "Celery worker consuming summary queue; runs nightly"
}
]
}Only include keys you can detect with reasonable confidence. Everything else stays for Clarify.
Keep the top-level model_provider compatibility field, but derive it from the primary unit's
model_source.provider. model_provider is not a scoring key. The structured per-unit
model_source object is authoritative for Model Recommend and the downstream handoff.
Step 3 — Tell the user what was detected
List the detected signals so the user can correct them in Clarify. These pre-fills let Clarify skip questions (Clarify asks fewer for build_deploy/migrate).
Determinism boundary (important): these detections are a best-effort LLM interpretation of code, NOT deterministic facts. They become inputs to the deterministic scoring engine, so a wrong detection silently biases scoring. Mitigation: (1) only write a signal you can detect with high confidence — when unsure, omit it and let Clarify ask; (2) always present detected signals to the user as "detected: X (correct me if wrong)" so they have a correction opportunity before scoring runs. This is the one point where LLM interpretation enters the otherwise deterministic pipeline.
Step 4 — Write state
Set phases.discover = completed.
Requirements and compatibility
Written for: Claude Code
Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta
“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.
Needs: Names Claude Code tools, Needs Python
Security
What our scanner noticed· Low risk
- Scripts that your agent may run
This skill includes scripts that your agent may run. Read them before you use it.
A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.
Attribution and licence
- Published by
- AWS
- Publisher
- AWS (Official)
- Source
- aws/agent-toolkit-for-aws · plugins/aws-startup-advisor/skills/agent-advisor · commit 0a218f8
- Licence
- Apache License 2.0 (redistribution allowed)
- LICENSE (added by AutoGineer from LICENSE)
- NOTICE (added by AutoGineer from NOTICE)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
AutoGineer is not affiliated with or endorsed by AWS or AWS.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download agent-advisor.zip