15 KiB
agent-claw: Final Architecture Design
Decisions locked 2026-05-04
System Overview
A serverless personal assistant running on AWS AgentCore. No always-on compute. Telegram-first, channel-agnostic by design. Two AgentCore runtimes: one user-facing, one background (REM).
Telegram ──→ API GW ──→ Lambda(tg-ingest) ──→ SQS FIFO ──→ Lambda(agent-runner)
│
InvokeAgentRuntime(Runtime 1)
│
Runtime 1: main assistant
reads: AgentCore Memory, factbase (Gateway),
S3 workspace
delivers via: channel adapter → Telegram
│
session ends (idle)
│
EventBridge ──→ Lambda(rem-runner)
│
InvokeAgentRuntime(Runtime 2)
│
Runtime 2: REM agent
calls: factbase workflow add/maintain
reads: AgentCore Memory session summary
writes: factbase (via AgentCore Gateway)
EventBridge ──→ Lambda(heartbeat-runner) ──→ InvokeAgentRuntime(Runtime 1, heartbeat prompt)
EventBridge ──→ Lambda(rem-runner) ──→ InvokeAgentRuntime(Runtime 2, nightly maintain)
CDK Stack
Stack: AgentClawStack
All resources are create-or-use-existing via CDK context parameters or SSM lookups. Deploy is idempotent.
// Context parameters (CDK --context flags or cdk.json)
// agentcore-memory-arn: existing memory ARN (optional, creates if absent)
// workspace-bucket-name: existing S3 bucket (optional, creates if absent)
// telegram-bot-token-secret-arn: existing secret (required before deploy)
// factbase-gateway-url: AgentCore Gateway endpoint for factbase tools
Resources provisioned:
| Resource | Type | Notes |
|---|---|---|
workspace-bucket |
S3 | SOUL.md, AGENTS.md, USER.md, HEARTBEAT.md |
session-store |
DynamoDB | actor_id → {session_id, created_at} |
personal-memory |
AgentCore Memory | SUMMARIZATION + USER_PREFERENCE strategies |
runtime-1 |
AgentCore Runtime | main assistant, CodeZip, Sonnet |
runtime-2 |
AgentCore Runtime | REM agent, CodeZip, Haiku |
message-queue |
SQS FIFO | MessageGroupId=actor_id, serializes per user |
rem-queue |
SQS | triggers REM after session idle |
tg-ingest |
Lambda | validates webhook, enqueues, sends typing |
agent-runner |
Lambda | SQS trigger, invokes Runtime 1, delivers response |
rem-runner |
Lambda | EventBridge + SQS trigger, invokes Runtime 2 |
webhook-api |
API GW HTTP | POST /telegram → tg-ingest |
heartbeat-rule |
EventBridge | every 30m → agent-runner (heartbeat prompt) |
rem-nightly |
EventBridge | nightly → rem-runner (factbase maintain) |
secrets |
Secrets Manager | bot-token, brave-api-key |
Runtime 1: Main Assistant
Language: Python + Strands + bedrock-agentcore SDK
Deploy: CodeZip
Model: us.anthropic.claude-3-7-sonnet-20250219-v1:0
Lifecycle: idleRuntimeSessionTimeout: 21600 (6hr), maxLifetime: 28800 (8hr)
Session start
- Check in-memory cache for workspace files (SOUL.md, AGENTS.md, USER.md)
- If cache miss: load from S3, cache in-memory
RetrieveMemoryRecords(query=recent_message, namespace=/preferences/actor_id/)→ top-k preferencesListEvents(session_id)→ recent conversation turns (if resuming)- Build system prompt: workspace files + retrieved memory
- Initialize Strands agent with
AgentCoreMemorySessionManager
Tools
| Tool | Description |
|---|---|
send_message(text, metadata?) |
Delivers via channel adapter (decoupled) |
web_search(query) |
Brave Search API |
web_fetch(url) |
HTTP + readability extraction |
read_workspace_file(path) |
S3 get_object |
write_workspace_file(path, content) |
S3 put_object + update in-memory cache |
search_memory(query) |
AgentCore Memory RetrieveMemoryRecords |
factbase_search(query) |
factbase via AgentCore Gateway |
NOT included in Runtime 1: factbase write tools (writes are async via Runtime 2 only)
Channel adapter (decoupled)
class ChannelAdapter(Protocol):
def send_message(self, text: str) -> str: ... # returns message_id
def send_typing(self) -> None: ...
def edit_message(self, msg_id: str, text: str) -> None: ...
class TelegramAdapter:
def send_message(self, text): ...
def send_typing(self): ... # sendChatAction(typing)
def edit_message(self, msg_id, text): ...
# Future: SlackAdapter, DiscordAdapter
# Channel type + target_id come from SQS message payload
# agent-runner instantiates correct adapter before invoking Runtime 1
# adapter is passed in payload, not hardcoded in agent
The agent calls send_message(text) — it doesn't know or care about Telegram specifics.
Runtime 2: REM Agent
Language: Python + Strands + bedrock-agentcore SDK
Deploy: CodeZip
Model: us.anthropic.claude-3-5-haiku-20241022-v1:0
Lifecycle: short sessions, maxLifetime: 1800 (30min)
Trigger: session ended
- agent-runner Lambda enqueues session summary to rem-queue when main session goes idle
- rem-runner Lambda invokes Runtime 2 with:
- session summary (from AgentCore Memory SUMMARIZATION)
- actor_id for namespace routing
Trigger: nightly (EventBridge cron)
- rem-runner Lambda invokes Runtime 2 with: "Run factbase maintain"
Tools
| Tool | Description |
|---|---|
factbase_workflow_add(summary) |
factbase workflow add via AgentCore Gateway |
factbase_workflow_maintain() |
factbase maintain workflow |
read_workspace_file(path) |
S3 (for HEARTBEAT.md context) |
write_workspace_file(path, content) |
S3 (update HEARTBEAT.md if needed) |
REM flow
Receive: session summary text
→ factbase workflow add: "Here is a conversation summary. Extract any new or updated
facts about people, projects, or decisions and add/update factbase entities."
→ factbase guides agent through: what's new, what conflicts, what to write
→ On nightly schedule: factbase workflow maintain
Lambda: tg-ingest
Trigger: API Gateway POST /telegram Runtime: Python 3.12
def handler(event, context):
# 1. Validate Telegram secret token header
if not validate_telegram_secret(event):
return {"statusCode": 403}
# 2. Parse Telegram Update
update = json.loads(event["body"])
chat_id = extract_chat_id(update)
message_text = extract_message(update)
# 3. Send typing action immediately (non-blocking)
send_telegram_typing(chat_id)
# 4. Enqueue to SQS FIFO
sqs.send_message(
QueueUrl=MESSAGE_QUEUE_URL,
MessageGroupId=str(chat_id), # serializes per conversation
MessageDeduplicationId=str(update["update_id"]),
MessageBody=json.dumps({
"channel": "telegram",
"chat_id": chat_id,
"messages": [{"text": message_text, "from": update["message"]["from"]}],
"update_id": update["update_id"],
"timestamp": update["message"]["date"],
})
)
return {"statusCode": 200} # Telegram acks within 3s requirement
Lambda: agent-runner
Trigger: SQS FIFO (batch_size=10, MessageGroupId=actor_id) Runtime: Python 3.12
def handler(event, context):
# 1. Batch all SQS records (same actor, serialized by FIFO)
records = [json.loads(r["body"]) for r in event["Records"]]
channel = records[0]["channel"]
chat_id = records[0]["chat_id"]
actor_id = f"{channel}:{chat_id}"
# 2. Look up or create session in DynamoDB
session_id = get_or_create_session(actor_id)
# 3. Bundle messages
if len(records) == 1:
prompt = records[0]["messages"][0]["text"]
else:
msgs = "\n".join(f"[{i+1}] {r['messages'][0]['text']}" for i, r in enumerate(records))
prompt = f"You have {len(records)} queued messages:\n{msgs}"
# 4. Invoke Runtime 1
payload = {
"prompt": prompt,
"actor_id": actor_id,
"session_id": session_id,
"channel_adapter": {"type": channel, "target_id": str(chat_id)},
"bot_token_secret_arn": BOT_TOKEN_SECRET_ARN,
}
response = agentcore.invoke_agent_runtime(
agentRuntimeArn=RUNTIME_1_ARN,
runtimeSessionId=session_id,
payload=json.dumps(payload).encode()
)
# 5. Stream response — agent delivers its own messages via send_message tool
for chunk in response["response"]:
pass # consume stream; agent handles delivery internally
# 6. Enqueue session summary for REM processing
sqs.send_message(QueueUrl=REM_QUEUE_URL, MessageBody=json.dumps({
"actor_id": actor_id,
"session_id": session_id,
}))
Lambda: rem-runner
Trigger: SQS (rem-queue) + EventBridge (nightly) Runtime: Python 3.12
def handler(event, context):
# Determine trigger source
if is_eventbridge(event):
# Nightly: run maintain on factbase
payload = {"prompt": "Run factbase maintain workflow to consolidate memories, detect conflicts, and generate review questions."}
session_id = f"rem-nightly-{today()}"
else:
# Post-session: extract new facts
record = json.loads(event["Records"][0]["body"])
actor_id = record["actor_id"]
session_id = record["session_id"]
summary = get_session_summary(actor_id, session_id)
payload = {
"prompt": f"Conversation summary to process:\n\n{summary}\n\nExtract new entities, facts, and updates.",
"actor_id": actor_id,
}
# Use dedicated REM session (separate from user session)
session_id = f"rem-{actor_id}-{today()}"
agentcore.invoke_agent_runtime(
agentRuntimeArn=RUNTIME_2_ARN,
runtimeSessionId=session_id,
payload=json.dumps(payload).encode()
)
Factbase KB Design (perspective.yaml)
organization: Daniel Levy
focus: Personal chief of staff KB — professional network, projects, commitments, decisions
entity_types:
person: Professional contacts — colleagues, customers, AWS team, interviewers, vendors
company: Organizations — AWS customers, target employers, partners, tools/services
project: Active and past projects — work deliverables, side projects, products
commitment: Explicit commitments made (by Daniel or to Daniel) with due dates
decision: Significant decisions made, rationale, outcomes
meeting: Key conversations — participants, decisions, follow-ups
review:
stale_days: 90
required_fields:
person: [current_role, company, last_contact_date]
commitment: [due_date, status, owner]
decision: [date, outcome]
S3 Workspace Layout
s3://agent-claw-workspace-{account_id}/
├── SOUL.md ← personality, tone, operating rules
├── AGENTS.md ← workspace conventions, memory rules
├── USER.md ← Daniel's profile, preferences
├── IDENTITY.md ← agent name, emoji
├── HEARTBEAT.md ← periodic task checklist
└── TOOLS.md ← tool-specific notes
Seeded from existing OpenClaw workspace files. CDK deploy script copies them if bucket is new.
Project Structure
agent-claw/
├── cdk/ ← CDK TypeScript stack
│ ├── bin/agent-claw.ts
│ ├── lib/agent-claw-stack.ts
│ └── lib/constructs/
│ ├── agentcore-runtime.ts
│ ├── memory.ts
│ └── eventbridge-rules.ts
├── src/
│ ├── runtime-1/ ← Main assistant (CodeZip deploy)
│ │ ├── main.py
│ │ ├── tools/
│ │ │ ├── __init__.py
│ │ │ ├── web.py ← web_search, web_fetch
│ │ │ ├── workspace.py ← read/write S3 workspace files
│ │ │ ├── memory.py ← search_memory (AgentCore Memory)
│ │ │ └── messaging.py ← send_message (channel adapter)
│ │ ├── channels/
│ │ │ ├── adapter.py ← ChannelAdapter protocol
│ │ │ └── telegram.py ← TelegramAdapter
│ │ ├── prompt_builder.py ← builds system prompt from workspace + memory
│ │ └── pyproject.toml
│ ├── runtime-2/ ← REM agent (CodeZip deploy)
│ │ ├── main.py
│ │ ├── tools/
│ │ │ ├── factbase.py ← factbase workflow tools via AgentCore Gateway
│ │ │ └── workspace.py ← shared workspace tools
│ │ └── pyproject.toml
│ └── lambdas/
│ ├── tg-ingest/
│ │ ├── handler.py
│ │ └── requirements.txt
│ ├── agent-runner/
│ │ ├── handler.py
│ │ └── requirements.txt
│ └── rem-runner/
│ ├── handler.py
│ └── requirements.txt
├── workspace/ ← seed files for S3 workspace
│ ├── SOUL.md
│ ├── AGENTS.md
│ ├── USER.md
│ └── HEARTBEAT.md
└── README.md
Build Phases
Phase 0 — Infrastructure spike (1-2 days)
- CDK stack: deploy API GW + tg-ingest Lambda + SQS + AgentCore Runtime 1 (minimal agent)
- Test: Telegram message → Lambda → SQS → Runtime 1 → response back to Telegram
- Confirm: cold start latency, session warm behavior, basic round-trip
Phase 1 — Full main assistant (1 week)
- Runtime 1 with all tools: web_search, web_fetch, workspace read/write, send_message
- System prompt builder loading from S3
- AgentCore Memory: short-term turns + USER_PREFERENCE
- Channel adapter: TelegramAdapter with typing action
- DynamoDB session mapping
Done when: full conversation works with personality, web search, and memory across messages.
Phase 2 — REM + factbase (1 week)
- Runtime 2 deployed with factbase workflow tools
- rem-runner Lambda + rem-queue
- Post-session trigger from agent-runner
- EventBridge nightly REM schedule
- factbase perspective.yaml defined, KB seeded
Done when: after a conversation, REM fires and writes relevant facts to factbase; next session the agent can search factbase for context.
Phase 3 — Heartbeat + scheduling (3-4 days)
- EventBridge heartbeat rule (every 30m → Runtime 1 with heartbeat prompt)
- HEARTBEAT.md checklist loaded at heartbeat time
- HEARTBEAT_OK suppression
- Dynamic cron via
create_crontool (EventBridge SDK)
Phase 4 — Polish
- Streaming progress (Telegram edit-in-place)
- Commitments system
- Error handling, DLQ, CloudWatch alarms
- Multi-user support (actor_id scoping already built in)
Design finalized 2026-05-04. All pre-development questions resolved.