sharaf/agentic-harness-architect
Design, build, or audit a coding agent, agentic loop, tool-use harness, or autonomous coding system — covering loop architecture, action space, context strategy, observation formatting, evaluation, error handling, prompt engineering, and task decomposition. Use when the user wants to design an agent, build a coding agent, scaffold an agentic system, architect a tool-use loop, review an existing agent harness for improvements, fix context bloat or compaction problems, tune observation formatting or tool output handling, debug agent loop or termination issues, design a system prompt or evaluator prompt for an agent, set up or redesign an agent evaluation pipeline, plan multi-agent orchestration, or specify how an agent should manage context, tools, prompts, evaluation, or recovery (greenfield design or audit mode).
100
100%
Does it follow best practices?
Impact
100%
1.23xAverage score across 4 eval scenarios
Passed
No known issues
Agentic Harness Architect
A Tessl skill for designing, building, or auditing agentic coding harnesses — the surrounding architecture (not the model) that determines an agent's quality ceiling.
What this skill does
Walks through the full architecture of a coding agent across 11 phases — loop design, action space, context management, evaluation, error handling, prompt engineering, decomposition, and simplification — and produces a structured design document. Two modes:
- greenfield — design a new harness from stated requirements
- audit — analyze an existing harness and propose targeted improvements
Every decision is grounded in benchmarks from 781 sources across 10 sub-domains (agent loop design, action space design, observation formatting, context window management, multi-agent orchestration, evaluation, prompt engineering, error handling, task decomposition, iterative simplification).
When to use
Trigger phrases that activate the skill:
- "Design an agent" / "build a coding agent" / "scaffold an agentic system"
- "Architect a tool-use loop" / "review an agent harness"
- "Fix context bloat or compaction problems"
- "Tune observation formatting" / "fix tool output handling"
- "Debug agent loop or termination issues"
- "Specify how an agent should manage context, tools, or recovery"
Install
tessl install sharaf/agentic-harness-architectOr pin a version:
tessl install sharaf/agentic-harness-architect@0.3.6How it's organized
| Layer | Where | Purpose |
|---|---|---|
| Entry point | SKILL.md | Quick Decision Triage + phase index + output format |
| Phase detail | references/phase-*.md | One reference file per architectural phase (1-10) |
| Guardrails | references/guardrails.md | Must-not-do rules across all phases |
| Decision flowcharts | references/decision-flowcharts.md | Architecture sizing, loop selection, context strategy, evaluation, decomposition, error recovery, simplification |
| Success criteria | references/success-criteria.md | Benchmarks + quality gates the design must hit |
The SKILL.md itself is intentionally short (~1.5K tokens) — progressive disclosure delegates depth to references, so the on-demand context cost stays modest.
Output format
The skill produces a structured design document. For greenfield mode, sections include Architecture Decision, Loop Design, Action Space, Observation Formatting, Context Management, Evaluation Design, Error Handling, Prompt Architecture, Decomposition Strategy, Complexity Budget, and Key Metrics. For audit mode, sections include Current Architecture Summary, Identified Issues, Improvement Sequence, Components to Remove, and Migration Path.
Each section includes the decision, the rationale citing benchmarks, alternatives considered, and conditions under which to revisit.
Benchmarks the skill targets
| Metric | Target |
|---|---|
| Tool definitions | <20% of context budget |
| Context utilization | 40-60% (FIC target) |
| Generator-Critic iterations | 2-3 max |
| Agent count | 2-4 (saturation threshold) |
| Decomposition depth | Max 3 levels |
| Error recovery rate | >70% |
Eval results
Latest published registry run tested claude:claude-sonnet-4-6 across 4 scenarios:
| Metric | Result |
|---|---|
| Baseline average | 81% |
| With-context average | 99% |
| Uplift | 1.22x |
The current bundle passes local Tessl lint and skill review at 100%.
- Workspace
- sharaf
- Visibility
- Public
- Created
- Last updated
- Publish Source
- CLI
- Badge
