How To Be A World-Class Agentic Engineer
A practitioner's guide to maximizing agent capabilities with minimal tooling — covering context management, sycophancy exploitation, task contracts, and iterative rules/skills architecture.
Overview
Most developers trying to get the most out of AI agents fall into a trap: accumulating frameworks, plugins, and harnesses in search of a magic combination. This article argues the opposite — that a barebones CLI setup (Claude Code or Codex) combined with a few fundamental principles will outperform any complex toolchain. The author draws from extensive production experience building agentic factories for signals, infrastructure, and data pipelines.
Less Is More
Foundation companies are on a generational run. Every new generation of agent intelligence changes optimal workflows, which means locking into external harnesses creates dependencies on solutions for problems that may soon disappear. If a capability is truly groundbreaking, frontier companies will incorporate it into their base products — skills, memory, subagents, and planning all started as external solutions before becoming core features. The practical advice: just update your CLI tool regularly and read the changelog.
Context Is Everything
The single most important principle is giving agents exactly the information they need and nothing more. Using many plugins and dependencies causes context bloat — agents become overwhelmed with irrelevant information. The key practice is separating research from implementation:
- Vague prompt: “Build an auth system” forces the agent to research alternatives, filling context with unnecessary options
- Precise prompt: “Implement JWT authentication with bcrypt-12 password hashing, refresh token rotation with 7-day expiry” lets the agent focus purely on execution
When you don’t know the implementation details, create a research task first, decide on the approach, then hand a fresh-context agent the specific implementation work.
Exploiting Sycophancy
Agents are engineered to please, which means biased prompts produce biased results. Saying “find me a bug” will make the agent engineer one if needed. Instead, use neutral prompts: “Search through the database, follow the logic of each component, and report all findings.”
The author describes a powerful adversarial pattern for bug-finding:
- A bug-finder agent scored on impact (incentivized to find everything, producing a superset of possible bugs)
- An adversarial agent scored for disproving bugs with penalties for wrong calls (producing a subset of real bugs)
- A referee agent that scores both, believing the author holds ground truth
This multi-agent pattern exploits each agent’s desire to please in opposing directions, producing high-fidelity results.
Task Contracts and Completion
Agents know how to start tasks but struggle with knowing when to stop. The solution is explicit contracts: define tests that must pass, screenshots that must verify design/behavior, and embed these into a {TASK}_CONTRACT.md. A stop-hook prevents session termination until all contract conditions are met.
For long-running automation, the author recommends one session per contract rather than marathon 24-hour sessions, which cause context bloat from unrelated work. An orchestration layer creates new contracts as needed and spawns fresh sessions for each.
Rules, Skills, and Iteration
Treat your CLAUDE.md as a logical, nested directory of where to find context — not a monolithic document. Use conditional branching: “If coding, read coding-rules.md; if tests are failing, read test-failing-rules.md.”
- Rules encode preferences (what not to do, how to behave)
- Skills encode recipes (specific approaches to specific problems)
Add rules and skills iteratively based on real friction. But as they accumulate, they’ll start contradicting each other and causing context bloat — so periodically consolidate and clean up. This cycle of iterate-then-consolidate is the core workflow.
Key Takeaway
Strip your agent setup to the bare minimum, obsessively manage context, and build up rules and skills iteratively from real usage — the simplicity is the superpower, not the toolchain.