Coordinate complex work
Break a goal into bounded tasks, route them to appropriate specialists, and keep responsibility clear across handoffs.
Models put to work
Agentic Engineering
Agents need a system around them.
Useful agents need more than instructions. They need the right context, bounded capabilities, durable task state, and a way to recover when something fails. My focus is the engineering around the model: how an agent gets work, what it can do, and how its actions are evaluated and governed.
Explore the capabilities- Harness engineering
- Multi-agent systems
- Context & memory
- Tool integration
- Evaluation & oversight
01
Problems worth solving
Connect knowledge to action
Give an agent relevant documents, system state, and tools without treating every available piece of information as useful context.
Make execution accountable
Track decisions, approvals, tool calls, and outcomes so a person can understand and control what happened.
02
Explore the capabilities
01Agent harnesses & durable execution
Design the runtime that carries a task from intake to completion. Explicit state and recoverable steps make interruption, retry, and human review part of the system.
Methods & tools
- Mission state
- Task routing
- Retries and checkpoints
- Execution logs
Example deliverables
- Harness architecture
- Task lifecycle
- Failure and recovery paths
02Multi-agent orchestration
Choose coordination patterns according to the work. Delegation, specialist review, and bounded collaboration should improve task quality enough to justify their added cost.
Methods & tools
- Managers and specialists
- Creator–verifier workflows
- Structured handoffs
- Bounded negotiation
Example deliverables
- Team definitions
- Delegation contracts
- Quality and cost comparisons
03Context engineering & memory
Separate immediate task context from durable knowledge and prior experience. Retrieve relevant evidence, preserve provenance, and decide deliberately what becomes persistent memory.
Methods & tools
- Retrieval
- Semantic search
- Session and episodic memory
- Context selection
Example deliverables
- Context assembly pipeline
- Memory boundaries
- Retrieval evaluation
04Tools, permissions & human oversight
Connect agents to useful capabilities through explicit interfaces and permissions. Consequential actions need clear authority, approvals, and an auditable link back to the originating task.
Methods & tools
- MCP integration
- Tool contracts
- Least privilege
- Approval gates
Example deliverables
- Tool registry
- Permission policy
- Approval and audit flow
05Agent evaluation & observability
Evaluate complete tasks rather than a single appealing response. Inspect tool use, completion quality, failure cases, latency, and cost across repeatable scenarios.
Methods & tools
- Task-based evaluations
- Trace inspection
- Regression scenarios
- Cost tracking
Example deliverables
- Evaluation scenarios
- Execution traces
- Operational acceptance criteria
03
From question to working system
- 01
Define the job
Identify the intended outcome, allowed actions, and the points where a human must remain in control.
- 02
Design the boundaries
Specify tools, context, identities, permissions, and task state.
- 03
Build a bounded workflow
Connect the runtime and integrations around a narrow, useful task.
- 04
Evaluate and extend
Test realistic success and failure cases before broadening the agent's responsibilities.
What work should your agents be able to own?
Start with one useful workflow, its tools, and the boundaries that make delegation acceptable.