AI-SRE / AgentOps
Measure token cost per incident, latency, quality and false positives through the internal Aurora pilot.
SPECIALIZATION CURRICULUM · AWS + GENERATIVE AI
Agentic
Operations
Build the team capability behind Infra Fox’s real wedge: AI-SRE / AgentOps, safe-agent deployment and GenAI FinOps—grounded in AWS, measured on internal systems and proven before client rollout.
Enter curriculum ↓10 modules · 16 weeks · evidence-gated
01 / POSITIONING TO CAPABILITY
The curriculum serves the strategy—not the other way around.
General AWS/DevOps remains the base-load. This curriculum develops only the capabilities needed to earn credibility in the Agentic Ops wedge.
AI security
Turn the OpenClaw work into a repeatable safe-agent deployment assessment.
GenAI FinOps
Extend existing cloud cost credibility into model and agent unit economics.
Deliberately parked
Broad GenAI app building, new verticals and unrelated AI requests that do not strengthen the wedge.
02 / NON-NEGOTIABLE FOUNDATION
Six mental models before autonomy.
Teach them in a spiral through the pilot. They are gates to safe judgment, not a six-week classroom prerequisite.
Agent fundamentals
LLMs, tokens, context windows, prompts, tools, agents, workflows, state, memory and human approval boundaries.
AWS runtime
IAM, VPC, Lambda/ECS, Step Functions, Bedrock, AgentCore, CloudWatch, KMS and Secrets Manager.
Reliability
Signals, failure modes, retries, timeouts, idempotency, queues, escalation, SLI/SLO and incident response.
Observability
Structured events, logs, metrics, OTEL traces, model/tool spans, latency, quality, false positives and attribution.
AI security
Prompt injection, tool abuse, excessive agency, data leakage, identity, sandboxing, guardrails and auditability.
AI economics
Tokens, model routing, cache, tool-call cost, unit economics, budgets, anomaly detection and cost per outcome.
03 / TARGET DEPTH
Different depth for different ownership.
The whole team reaches L2 across the map. Builders reach L3. The pod lead and Suresh reach L4 in the wedge; L5 follows real client evidence.
Aware
Describe the problem and place the capability on the Agentic Ops map.
Understand
Explain architecture, data flow, trust boundaries, cost and failure modes.
Practitioner
Build and troubleshoot with docs/AI; produce evidence and a runbook.
Independent
Design production trade-offs, lead incidents and review others’ work.
SME
Advise clients, define standards, package offers and mentor the team.
WHOLE TEAMFoundations L2 · Bedrock L2 · Security L2 · Observability L2
AI OPS PODAgentCore L3 · LangGraph L3 · OTEL L4 · Security L3 · FinOps L4
LEADERSHIPArchitecture L4 · Client discovery L4 · Packaging L4 · Governance L4
04 / PROJECT LADDER
One internal system. Ten evidence-gated modules.
Each module uses Problem → Concepts → Build → Explain → Verify → Break → Reflect. Nothing is complete until its gate is passed.
01ORIENT
Agentic systems map
Trace user intent → model → tool → AWS resource → observation → approval → outcome.
CONCEPTS
Agents vs workflows · autonomy · state · tool contracts · human-in-the-loop
BUILD
Diagram the proposed AI-SRE loop and threat boundaries before writing code.
BREAK / INCIDENT
Given an unsafe tool call, identify which boundary failed.
02FOUNDATION
Bedrock & model economics
Select models using quality, latency, context and cost—not popularity.
CONCEPTS
Bedrock · inference · token accounting · model routing · quotas · privacy
BUILD
Create a small evaluation harness comparing two models on ops tasks.
BREAK / INCIDENT
Diagnose throttling, runaway tokens and degraded answer quality.
03BUILD
Deterministic orchestration
Know when Step Functions is safer than an open-ended agent loop.
CONCEPTS
State machines · retries · timeouts · idempotency · compensation · approvals
BUILD
Orchestrate a read-only diagnostic workflow with explicit approval gates.
BREAK / INCIDENT
Recover from duplicate execution and partial tool failure.
04BUILD
LangGraph agent workflow
Build bounded, inspectable agent behavior with explicit state and exits.
CONCEPTS
Graphs · nodes · edges · state · memory · tool calling · termination
BUILD
Implement symptom → evidence → hypothesis → recommendation with no write access.
BREAK / INCIDENT
Stop a looping agent and prove the termination safeguard.
05PLATFORM
Bedrock AgentCore runtime
Deploy an agent with durable identity, isolation, observability and lifecycle ownership.
CONCEPTS
Runtime · identity · gateway/tools · memory · policy · environment separation
BUILD
Deploy the internal Aurora pilot in a sandboxed AWS environment.
BREAK / INCIDENT
Repair missing permissions without broadening access unnecessarily.
06OBSERVE
GenAI observability with OTEL
See one request across model, tool, infrastructure and business outcome.
CONCEPTS
Traces · spans · correlation · CloudWatch · latency · quality · false positives
BUILD
Instrument model calls, tool calls and approval waits; create an ops dashboard.
BREAK / INCIDENT
Find a latency regression using traces rather than guesses.
07CONTROL
Token cost per incident
Turn LLM usage into an operational unit metric suitable for pricing and governance.
CONCEPTS
Attribution · unit cost · budgets · anomaly detection · showback · optimization
BUILD
Create the reusable Agent Cost Governance dashboard.
BREAK / INCIDENT
Investigate a 3× cost spike and distinguish usage, model and retry causes.
08SECURE
Safe-agent deployment
Design least-agency systems that resist hostile input and contain tool misuse.
CONCEPTS
Prompt injection · Bedrock Guardrails · IAM · data boundaries · egress · audit
BUILD
Convert OpenClaw findings into an assessment checklist and test sandbox.
BREAK / INCIDENT
Respond to a prompt-injection attempt that reaches a privileged tool.
09KNOWLEDGE
Internal RAG assistant
Ground answers in owned runbooks while enforcing authorization and citations.
CONCEPTS
Knowledge Bases · chunking · retrieval · metadata · access control · evaluation
BUILD
Create an assistant over approved Infra Fox runbooks and internal documents.
BREAK / INCIDENT
Diagnose a confident answer sourced from stale or unauthorized content.
10VALIDATE
AI-SRE readiness audit
Translate internal evidence into one safe, repeatable client engagement.
CONCEPTS
Discovery · maturity · risk · architecture · economics · roadmap · boundaries
BUILD
Run the audit with one warm design partner after internal metrics exist.
BREAK / INCIDENT
Handle a no-access or low-data client environment without inventing certainty.
05 / CAPSTONE SYSTEM
The internal Aurora / AgentCore pilot.
The pilot is the learning environment, evidence generator and future demo asset. It runs against Infra Fox’s own environment before any client system.
Operational signalCloudWatch / alert / runbook→
Agent loopObserve · reason · retrieve · propose→
Approval boundaryHuman reviews all write actions→
Bounded toolLeast-privilege AWS operation→
EvidenceTrace · outcome · latency · tokens · cost
per incident investigated
time to useful recommendation
accepted / rejected recommendations
unhelpful actions or escalations
blocked attacks and policy violations
engineer time saved per incident
SEQUENCING RULE
Do not skip to a client pilot. First establish the internal loop and cost-governance dashboard; only then offer one AI-SRE Readiness Audit to a warm design partner.
06 / TASK MIX
Build 40%. Incident 35%. Design 25%.
Because agentic systems fail in new ways, this specialization deliberately gives more weight to troubleshooting and safety review.
Instrument one agent run end to end
- Problem: no reliable cost attribution
- Concepts: spans, tokens, correlation
- Build: OTEL + CloudWatch pipeline
- Verify: reconcile trace and bill
- Break: lose one tool span
- Reflect: improve the schema
The agent recommends the wrong action
- Start with symptom and impact
- Inspect prompt, retrieval and tool traces
- Test competing hypotheses
- Isolate root cause
- Contain, fix and verify
- Write prevention action
Safe client deployment
- Draw trust and data boundaries
- Choose workflow vs agent
- Set autonomy and approval level
- Define identity and tool permissions
- Model cost and failure modes
- Defend the trade-offs
07 / DEFINITION OF DONE & REVIEW
A demo is not proof.
Every project or module ends with evidence, an architecture review, a failure exercise and a reusable artifact.
Architecture
Can the engineer draw model, tools, data, identities, networks, state and approvals?
1 2 3 4
Reliability
Are timeouts, retries, idempotency, fallbacks, failure domains and escalation explicit?
1 2 3 4
Security
Are prompt injection, excessive agency, secrets, data exposure and audit risks tested?
1 2 3 4
Evidence
Are quality, latency, false-positive and cost claims supported by reproducible measurements?
1 2 3 4
Economics
Is token cost tied to an incident or business outcome, with budgets and ownership?
1 2 3 4
Operations
Can someone else deploy, observe, stop, recover and roll back using the runbook?
1 2 3 4
AI ownership
Can the submitter explain, verify and modify every AI-generated artifact?
1 2 3 4
Client judgment
Are constraints, uncertainty, trade-offs and recommendations communicated honestly?
1 2 3 4
Agentic Ops Definition of Done
08 / TEAM MODEL
A two-person pod builds. The whole team compounds.
AI Ops Pod · 2 people
Own modules 03–08 and the Aurora pilot. One builder, one reviewer; rotate after the first validated milestone.
Team · 4 people
Complete foundation modules, shadow milestones, reproduce incidents and join weekly brown-bags.
Suresh
Own positioning, architecture gates, design-partner decision and strategic distillation into public content.
Everyone
Capture decisions, protect client data, follow approval gates and teach one verified lesson back to the team.
BUILD→REVIEW→CAPTURE→BROWN-BAG→SHADOW→ROTATE
09 / LEARNING & CREDENTIALS
Certification supports delivery. It does not replace it.
AWS Certified AI Practitioner
Whole team · baseline now
Target: L1–L2 vocabulary and AWS AI mapMachine Learning Engineer — Associate
Pilot builder · during pilot
Target: supporting ML delivery and operations depthGenerative AI Developer — Professional
Suresh + one lead · after pilot begins
Target: architecture and delivery depth backed by pilot evidenceInternal evidence—builds, incidents, reviews and teach-backs—sets the L-level. A certificate is supporting evidence only.
10 / 16-WEEK ROLLOUT
Learning, proof and positioning move together.
FOUNDATION + FIRST BUILD
Form the pod and make the system visible.
Baseline the team · complete modules 01–03 · start Aurora pilot · publish the already-earned OpenClaw security article · first brown-bag.
Gate: architecture approved and one read-only diagnostic flow running.11 / KNOWLEDGE & CONTENT LOOP
No content before the work exists.
Public content is the final stage of knowledge capture—not the starting point. Real numbers, no client names and no AI slop remain non-negotiable.
POC / incident→engineering note→reviewed runbook→internal teach-back→case study / blog→LinkedIn distillation
Aurora pilot findings and OpenClaw assessment work · every 4–6 weeks.
Token-cost dashboards and unit economics in the proven case-study format · every 4–6 weeks.
Continue useful AWS FinOps and security content from existing client work at the current cadence.