---
title: "Agentic Development: The Four Parts of a Delegation Loop"
description: "Agentic development turns AI assistance into delegated work: a harness runs the agents, skills supply domain knowledge, a sandbox bounds execution, and evals verify the result."
---

# Agentic Development: The Four Parts of a Delegation Loop

**Type:** [What is it?](https://yomu.fyi/visual-explainers/category/what-is-it.md)

![A diagram with a central Delegation Loop hub surrounded by four panels labeled Harness, Skills, Sandbox, and Verification, connected by clockwise arrows, with a bottom strip showing agent adoption and trust statistics.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/cfadd3eb-92ea-4528-8535-222bb3fdeecb/image-469b42cb-cb98-40de-84f2-574111aa1b27.webp)

> Agentic development turns AI assistance into delegated work: a harness runs the agents, skills supply domain knowledge, a sandbox bounds execution, and evals verify the result.

## Explanation

Agentic development is the shift from asking an AI for suggestions to delegating whole tasks to agents that run, act, and produce artifacts. Its essential parts are a harness that launches and coordinates agent sessions, skills that give an agent verified domain knowledge, a sandbox that separates trusted code from untrusted agent code, and a verification layer that tests results before they ship. Engineers care because agent adoption is climbing faster than trust in the output, so the parts built around the model decide whether delegation stays safe and repeatable.

## FAQ

### Does a more elaborate agent harness make an agent more capable?

Not necessarily. Under an equal time budget and the same frontier model, open-source state-of-the-art harnesses showed no advantage over a single session of a minimal coding agent using direct read, write, and bash primitives. The backbone model, not the surrounding machinery, was the primary performance driver.

### Why do agent skills need their own tests?

Skills silently misfire when their trigger descriptions are passive or too general, so the agent never routes to them. Eval loops generate positive and negative prompts to tune those descriptions, and frameworks like promptfoo turn the ephemeral tests into committed configs with a skill-used assertion that runs in CI. Domain skills with verified API signatures measurably cut correction cycles.

### Where should the trust boundary sit between an agent and its tools?

Running the agent harness and its tool calls in the same place creates a risky combination of private data, untrusted content, and external communication. Sidecars put trusted work such as credential injection and proxy traffic in an isolated container on the same host, which is faster than a remote service and keeps the agent out of it.

**Tags:** [Agentic Development](https://yomu.fyi/topic/agentic-development.md), [AI Agents](https://yomu.fyi/topic/ai-agents.md), [Coding Agents](https://yomu.fyi/topic/coding-agents.md), [Evaluation](https://yomu.fyi/topic/evaluation.md), [Human-in-the-Loop](https://yomu.fyi/topic/human-in-the-loop.md), [Sandboxing](https://yomu.fyi/topic/sandboxing.md)

## Related reading

### [Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills](https://yomu.fyi/post/build-applications-on-nvidia-bluefield-faster-with-nvidia-doca-agent-s.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Oct 1, 2026
- Type: Problem & solution

General-purpose AI agents lack domain-specific knowledge for NVIDIA DOCA development, often inventing APIs or skipping hardware checks, leading to correction cycles. DOCA AI agent skills provide a structured, machine-readable specification in a SKILL.md file containing verified API signatures, hardware capability requirements, and build constraints for components like Flow, GPUNetIO, and PCC. In a 65-prompt evaluation, agents without skills satisfied only 19% of graded checklist items, while agents with skills satisfied 100% across all prompts. The skills reduce correction cycles; in a side-by-side demo, the with-skills agent used 73% less handwritten code and 46% fewer hardware commands. The skills are available on GitHub and enable developers to build faster, ship more stable code, and deploy with fewer unknowns on BlueField infrastructure.


### [A look back before we look forward: A Developer Survey retrospective](https://yomu.fyi/post/a-look-back-before-we-look-forward-a-developer-survey-retrospective.md)
- Company: [Stack Overflow](https://yomu.fyi/company/stack-overflow.md)
- Author: Erin Yepis
- Published: Oct 1, 2026
- Type: Explainer

A retrospective analysis of the Stack Overflow Developer Survey compares 2024 and 2025 data across AI adoption, work settings, and demographic shifts. AI tool usage among developers rose from 62% to 79% year-over-year, while agent adoption jumped from 31% in 2025 to 59% in an April 2026 pulse survey, with ChatGPT and GitHub Copilot as primary gateways. Enthusiasm for AI tools cooled: favorable sentiment fell from 72.0% to 59.7% overall and from 72.2% to 52.8% among learners, and trust in accuracy declined from 49.1% to 37.4% for students, while skepticism among learners quadrupled from 6.4% to 26.3%. Fully remote work dropped from 38% to 32.4%, but 82% of respondents had some remote flexibility. The article identifies growing reliance on AI alongside skepticism and frames unanswered questions about learning, trust, and job satisfaction for the upcoming 2026 survey.


### [Master your skills: Building skills you can trust](https://yomu.fyi/post/master-your-skills-building-skills-you-can-trust.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Dejan Bosanac
- Published: Oct 1, 2026
- Type: Explainer

Skills in AI agent projects silently misfire when their trigger descriptions are passive or too general, causing the agent to skip using them. The authoring tool skill-creator, from Anthropic’s skills collection, provides an eval loop that generates positive and negative test prompts and iterates on description accuracy using a training and held-out test set to avoid overfitting. These tests are ephemeral, so the open-source framework promptfoo extends them into permanent configuration files that can be committed alongside skills. Promptfoo’s agent SDK providers test through the actual agent runtime and include a skill-used assertion to confirm proper routing, and its YAML-based configuration fits naturally into version control. The final step is wiring promptfoo evals into CI, using GitHub Actions with path filters to run on pull requests that modify skills, catching regressions before they ship and costing roughly $0.10-0.15 per eval.


### [A New Agentic Experience: JetBrains Air in IDEs – EAP Now Open](https://yomu.fyi/post/a-new-agentic-experience-jetbrains-air-in-ides-eap-now-open.md)
- Company: [Jetbrains](https://yomu.fyi/company/jetbrains.md)
- Author: Dominique Rolink
- Published: Oct 1, 2026
- Type: Announcement

JetBrains launched the Early Access Program for Air, an agentic experience in its IDEs available as a plugin or in 2026.3 EAP builds. Air orchestrates multiple AI agents in parallel, moving beyond the traditional AI chat model that JetBrains found insufficient for concurrent tasks. It acts as a conduit for existing agents like Codex, GitHub Copilot, Junie, and Claude, requiring no JetBrains AI subscription. The agentic-first UX organizes work into sessions with cost visibility, temporary worktrees, and local or cloud runs, with free Junie Lite for everyday tasks. Air ships with no agents installed, detects agents on the machine, and sends data directly to the user’s provider, while AI Assistant remains available as a separate plugin.


### [Sidecars: A low-latency trust boundary for Sandboxes](https://yomu.fyi/post/sidecars-a-low-latency-trust-boundary-for-sandboxes.md)
- Company: [Modal](https://yomu.fyi/company/modal.md)
- Published: Oct 1, 2026
- Type: Announcement

Modal introduces Sidecars, isolated containers that run alongside a main Sandbox on the same host, establishing a trust boundary between trusted and untrusted code. Sidecars enable communication across trust boundaries at least 3x faster than using separate Sandboxes, reducing latency for operation-heavy agent workloads. The feature addresses the security risk of running an agent’s harness and tool calls together, which creates the “lethal trifecta” of private data, untrusted content, and external communication. Existing solutions either move the harness to a remote service (adding network latency) or rely on egress controls that lack flexibility. Sidecars allow trusted services such as credential injection, proxy traffic, and harness logic to run locally alongside the Sandbox, with independent network policies and dynamic creation via the Modal SDK.


### [How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?](https://yomu.fyi/post/how-much-of-a-harness-does-a-strong-agent-need-for-autonomous-ml-engin.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Oct 1, 2026
- Type: Problem & solution

Recent autonomous machine learning engineering agents increasingly rely on elaborate machinery such as multi-agent orchestrators and dedicated retrieval subagents, motivated by progress stagnation over long-horizon cycles and limited Large Language Model primitives. This paper finds that under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline that uses direct read, write, and bash primitives. The authors argue via large-scale systematic ablation studies that the machinery layers become redundant in the coding agent setting. The backbone, not the harness complexity, is the primary driver for performance. Consequently, the effort spent elaborating hand-crafted harnesses around strong models yields poor returns for current MLE benchmarks.


### [How to scale agentic applications without creating AI sprawl](https://yomu.fyi/post/how-to-scale-agentic-applications-without-creating-ai-sprawl.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Julia Powell, Sirui Sun
- Published: Sep 30, 2026
- Type: Problem & solution

As agentic applications move from answering questions to taking actions, enterprises risk AI sprawl from duplicated integrations, inconsistent policies, and fragmented context. Databricks asserts scaling requires three pillars: choice across models and tools, context grounded in governed enterprise data and business semantics, and control through consistent permissions, policies, observability, and cost management. Its platform provides Agent Bricks for unified agent fleet building and governance, Omnigent as a common layer above agent harnesses, and Unity Gateway as a centralized control plane for models, agents, and tools. Context is shared via Unity Catalog for data and AI asset governance and Genie Ontology for common business concepts, letting teams reuse governed context across applications. Control uses Unity Gateway for access policies, budgets, and traces and Databricks Sandbox for isolated execution, with MLflow enabling evaluation and workflow inspection to maintain reliability as agents scale.


### [How Deterministic Controls Turn AI Output Into Reliable Prompt Templates](https://yomu.fyi/post/how-deterministic-controls-turn-ai-output-into-reliable-prompt-templat.md)
- Company: [Salesforce](https://yomu.fyi/company/salesforce.md)
- Author: Scott Nyberg
- Published: Sep 30, 2026
- Type: Explainer

Salesforce’s Assistive Authoring aimed to reduce prompt-template creation from 25 minutes to under a minute by using Agent Graph. A router node dispatches natural-language requests to specialized sub-agents for creation, refinement, or clarification while state variables carry context. Deterministic controls prevent the LLM from corrupting routing, record identity, or structured output through guarded transitions and a server-side transpose layer. A discovery service normalizes available data fields into an allowlist, and the LLM can only select from validated references. A human-in-the-loop gate ensures the agent never saves or executes, presenting output as an accept-or-reject diff only.