Agentic engineering for the whole team.
Training that runs from the agent loop to fine-tuning, tailored to your team: modules, depth and exercises chosen from your stack, your systems and your goals.

What the program covers.
- Harness engineering
- Context engineering
- MCP 2026-07-28 (stateless)
- Agent Skills (SKILL.md)
- AGENTS.md
- Hooks and subagents
- Tool search
- Code mode
- MCP Apps
- OAuth 2.1
- A2A
- AG-UI
- Prompt caching
- Context compaction
- Contextual retrieval
- Hybrid search + reranking
- ColPali
- GraphRAG
- Durable execution
- LLM-as-judge
- pass^k
- Trajectory evals
- Lethal trifecta
- CaMeL
- Egress allowlists
- OWASP Top 10 for Agentic Applications
- OpenTelemetry GenAI
- Model cascades
- AI gateway
- LoRA / QLoRA
- DPO / KTO
- RFT with GRPO
- Multi-LoRA serving
- Spec-driven development
- EARS requirements
- Property-based testing
- EU AI Act
- ISO/IEC 42001
- NIST AI RMF
Curriculum.
Harness fundamentals
Your team writes an agent loop from scratch and builds the repo harness: an AGENTS.md with commands and boundaries, skills for repeated playbooks, and hooks that block dangerous commands and require passing tests to finish.
- Agent loop with native tool calling
- AGENTS.md as a map of the repo
- Skills, hooks or subagents: which to use
- Verifying with tests, logs and a browser
- Revisiting the harness on each new model
Context engineering
Your team measures how much of the window goes to tools, instructions and history, orders the prompt to get cache hits, and decides when to compact, when to reset with a handoff file, and what to keep as memory.
- Context budget and context rot
- Progressive disclosure and tool search
- Prompt caching and stable prefixes
- Compaction, context editing and resets
- Agent memory and memory poisoning
RAG in production
Before indexing, your team checks if the corpus fits in the prompt. Then it builds hybrid search with a reranker, applies ACLs at query time, and measures recall@k on real questions for each chunking or embedding change.
- Parsing, chunking, contextual retrieval
- Hybrid search, reranking and recall@k
- Search as an agent tool
- GraphRAG and multimodal RAG: when to use
- ACLs, re-indexing and RAG poisoning
Tools and protocols
Your team designs workflow-level tools, deploys a stateless MCP server with OAuth, and picks the protocol by boundary: MCP for tools, A2A for another team's agents, AG-UI or MCP Apps for the interface.
- Tool design and code mode
- Stateless MCP servers with OAuth 2.1
- A2A: delegating to another team's agents
- AG-UI, MCP Apps and human approval
- Computer use as a last resort
Multi-agent orchestration
Multi-agent setups use 3 to 10 times more tokens than a single agent. Your team uses evals to check when that pays off, splits work by context, keeps writes in one agent, and runs long flows with checkpoints.
- When multi-agent pays off
- Orchestrator and independent verifier
- Workflows in code with typed handoffs
- Durable execution: checkpoints, retries
- Parallel fan-out, single-threaded writes
Evals
Your team builds a suite of 20 to 50 tasks drawn from real failures, calibrates the LLM judge against human labels, and blocks prompt, tool or model changes in CI when results drop beyond measured noise.
- Error analysis on real traces
- Graders: code, LLM-as-judge, human
- pass@k, pass^k and confidence intervals
- Outcome and trajectory evals
- Eval gates in CI and online evals
Guardrails and security
Adaptive attacks bypass most published prompt injection defenses. Your team threat-models each agent, isolates execution with a sandbox and a network allowlist, audits MCP servers, and runs red-team tests in CI.
- Lethal trifecta and Rule of Two
- Secure patterns: dual LLM and CaMeL
- Sandboxing, egress and permissions
- MCP, skill and plugin supply chain
- OWASP Top 10 for LLMs and agents (2026)
AgentOps: observability and cost
Your team instruments agents with OpenTelemetry, tracks the cost of each completed task by team and feature, caps tokens per run, and adopts model routing only if it beats the best single model on your own evals.
- Tracing with OpenTelemetry GenAI
- Cost per task and token budgets
- Routing, cascades and advisor models
- AI gateway: keys, limits and MCP
- SLOs, loop detection and kill switch
Model customization
Your team fine-tunes only if context engineering, RAG and automated prompt optimization fall short on evals. Then it trains a LoRA on an open-weight model, serves the adapters, and reruns safety regressions on every version.
- Fine-tune or not: an eval-first call
- LoRA and QLoRA SFT on open weights
- Preference tuning with DPO and KTO
- RFT and GRPO with verifiable rewards
- Distillation and multi-LoRA serving
Spec-driven development
Your team decides when a change needs a spec, writes acceptance criteria a test can check, reviews agent PRs with test evidence, and sets when an AI approval counts toward merge.
- When to write a spec and when to skip
- EARS criteria and property-based tests
- Specs in the repo and drift control
- Reviewing agent PRs with evidence
- Issue to merge: flow and metrics
AI governance
Your team classifies each system and agent under the EU AI Act with post-Omnibus deadlines, maps legal basis and data transfers (LGPD and GDPR) for each LLM flow, and writes per-team usage policies enforced in the tools.
- EU AI Act: deadlines and risk tiers
- LGPD, GDPR and cross-border transfers
- ISO/IEC 42001 mapped to NIST AI RMF
- AI inventory and an owner per agent
- Usage policy and documented AI literacy
Formats.
1 day
Executive workshop
For technical and product leadership. Where AI pays off, where it adds risk and how to decide.
3 days
Technical immersion
For engineers. Harness, MCP, evals and orchestration applied to a real repository.
8 weeks
Full program
The team, from developers to leadership, builds the company harness with our mentoring and puts it into production.
We shape the track around your stack.
A quick team assessment, the right modules and a real project as the final deliverable.