Coding Agent Harness / Agentic Developer Tools / Context Engineering
I work on the layer between LLMs and real developer workflows: execution control, permission boundaries, observable hooks, and cache-aware context design.
My current target is agent-harness engineering for coding agents: making model behavior controllable, inspectable, recoverable, and efficient enough for real developer loops.
For me, the Model reasons and generates; the Harness owns context assembly,
tool execution, permission control, state transitions, error recovery, traces,
and human intervention.
- Building and studying
Coding Agent Harnesssystems: planner/executor loops, human approval, pause/resume correctness, andPermissionRequestboundaries. - Designing hook surfaces for agent observability:
message_submit,turn_end,subagent_spawn,subagent_complete, traces, and tool-call attribution. - Practicing
Context Engineering: stable-prefix prompts,KV Cache,Prompt Cache, long-running agent state, and reusable agent skills/workflows.
| Capability | Engineering surface |
|---|---|
| Agent control plane | Planner/executor state, approval gates, resumable execution, rollback paths. |
| Observability hooks | Lifecycle events, traces, tool-call attribution, cost and failure-loop signals. |
| Context systems | Stable prompt prefixes, cache-aware state layout, reusable skills, durable notes. |
| Developer tools | CLI workflows, browser utilities, onboarding docs, reviewable OSS patches. |
| Area | Public evidence | What it shows |
|---|---|---|
| Permission and approval control | DeepSeek-Reasonix PR #3855PermissionRequest hook; PR #3935 approved-plan pause/resume fix | Harness control flow: permission gates, human approval, and resumable execution state. |
| Lifecycle and observability hooks | CodeWhale PR #2586 merged subagent lifecycle hooks; related hook work in #2578 and #2318 | Traceable agent events, extension points, and operational visibility. |
| Developer onboarding | CodeWhale PR #2278 first-run guide | I document agent tools so users can reach the first successful run faster. |
| Repository | Role |
|---|---|
| AresNing/DeepSeek-Reasonix | Fork for DeepSeek-native terminal coding-agent experiments and upstream contribution work. |
| AresNing/CodeWhale | Fork for coding-agent lifecycle hook contributions and terminal-agent workflow exploration. |
| AresNing/http-request-response-copy-helper | Browser-side developer utility; small, practical tooling for repeated engineering workflows. |
| AresNing.github.io | Personal writing archive and technical notes. |
- Start with DeepSeek-Reasonix PR #3855 for permission-boundary thinking.
- Then read CodeWhale PR #2586 for lifecycle hook design.
- Use the related PRs above as supporting evidence for state-machine and observability work.
- Reduce agent ambiguity into explicit state machines, durable events, and testable permission boundaries.
- Prefer small PRs with concrete behavior, reviewable diffs, and clear rollback paths.
- Treat prompts, caches, traces, and tool results as production interfaces, not incidental strings.
- Keep reusable knowledge in structured notes so future agents can pick up the same context without re-discovery.
Agent Harness · Coding Agent · Context Engineering · KV Cache ·
Prompt Cache · Tool Calling · Trace · CLI Tooling · Go ·
TypeScript · React