Part 9: Frontier Practices
Frontier practice is not a list of fashionable features. It is the discipline of connecting research, computer use, coding, background execution, model routing, and a durable local harness without losing safety or verifiability.
Chapter List
| Chapter | Topic | Core question |
|---|---|---|
| 27 | Deep Research | How do we gather, compare, cite, and synthesize evidence? |
| 28 | Computer Use | How do Agents operate interfaces safely and verify outcomes? |
| 29 | Agentic Coding | How do Agents edit code while preserving review and rollback boundaries? |
| 30 | Background Agents | How do long-running tasks survive cancellation, restart, and delayed events? |
| 31 | Tiered Model Strategy | How do we route by measured capability, cost, and latency? |
| 32 | Agent Harness | What belongs in the reliable execution loop around a model? |
| 33 | Building on the Harness — Kocoro | How do public OSS runtime concepts become a testable product platform? |
What Each Chapter Adds
Chapter 27: Deep Research
Evidence collection needs source quality, claim-to-citation mapping, deduplication, conflict handling, and a clear stop condition. More search is not automatically better research.
Chapter 28: Computer Use
A safe computer-use loop separates perception, proposed action, approval, execution, and postcondition verification. Visual success must be checked against application state whenever possible.
Chapter 29: Agentic Coding
A coding Agent needs a scoped worktree, explicit edit authority, reversible changes, tests, and a reviewable diff. Passing output is evidence, not permission to publish.
Chapter 30: Background Agents
Long-running work needs cancellation, checkpointing, resumability, idempotent tool calls, bounded retries, and visible terminal states. “Still running” is not a completion condition.
Chapter 31: Tiered Model Strategy
Model tiers are workload-specific policies, not fixed vendor names or quotas. Evaluate routing with task success, escalation rate, latency, and total cost—not with a single benchmark.
Chapter 32: Agent Harness
The harness owns context pressure, permissions, hooks, loop detection, progress reporting, and graceful termination. Exact thresholds and detector counts are implementation snapshots; the invariants matter more.
Chapter 33: Building on the Harness — Kocoro
Kocoro’s public engine and daemon illustrate named Agents, sessions, memory, scoped tools, MCP, scheduling, and local service integration. The native Kocoro Desktop product is separate and closed-source.
Learning Objectives
After this Part, you should be able to:
- design an evidence-backed research workflow;
- separate proposed UI actions from verified outcomes;
- define review, edit, commit, and publish as distinct coding gates;
- make long-running tasks cancellable, resumable, and idempotent;
- measure model routing instead of relying on fixed percentages;
- evaluate an Agent harness through contracts, traces, failure injection, and regression suites;
- distinguish public OSS implementation evidence from product-specific or private details.
Cost Optimization Without a Stale Price Table
Use a workload equation rather than memorizing a headline percentage:
monthly cost = sum(request count × input tokens × input rate + output tokens × output rate)
Record the pricing date, cache behavior, batch discounts, long-context premiums, and retry/escalation costs. Re-run the calculation whenever a model or provider changes.
Recommended Reading Paths
- Research builder: 27 → 31 → 32 → 33
- Coding Agent builder: 29 → 25 → 30 → 32 → 33
- Computer-use builder: 28 → 25 → 30 → 33
- Platform evaluator: 30 → 31 → 32 → 33
Public-source boundary: Shannon and the Kocoro engine/daemon may be used as public references. Examples in this book should remain conceptual or link to public OSS. Private repositories, production configuration, incidents, customer data, measured internal thresholds, and proprietary assets are out of scope.