Evaluating prompt design, instruction density, and context overhead across Claude Code, DeepSeek Coder, OpenAI Codex, OpenCode, T3 Code, and the Fleet Multi-Agent Harness with live empirical test telemetry.
fleet-prompt-bench
To prevent prompt optimization from becoming guesswork, the fleet uses an automated, reproducible benchmark harness (bin/fleet-prompt-bench) that executes A/B test runs across twin worktrees. Results are logged to an append-only ledger (~/.fleet/benchmarks/prompt-evals.jsonl).
Standard feature addition requiring handler dispatch, test mapping update in tools/fast-gate.sh, and .fleet/coverage-contract.json sync. Model: deepseek/deepseek-flash.
prompt-evals.jsonl)As language models scale in raw coding capability, the system prompt is no longer just a set of instructions; it is the operating system ABI (Application Binary Interface) that governs how models manipulate the host filesystem, handle fatal errors, manage context windows, and interact with other autonomous processes.
In this audit, we examine the production system prompts and control architectures of five premier coding agent implementations, benchmark them against the Exe.dev Fleet prompt architecture (roles/common.md, roles/coord.md, and role specifications), and identify four concrete improvements that reduce token bloat while boosting reasoning accuracy.
| Agent Runtime | Prompt Paradigm | Primary Strengths | Key Vulnerabilities / Gaps |
|---|---|---|---|
| Claude Code Anthropic |
Dynamic Modular Fragments | Clickable file:line hyperlinks; turn-by-turn <system-reminder> injection; strong distinction between exploration and destructive execution. |
Tightly bound to single interactive sessions; lacks distributed multi-agent ledger synchronization. |
| DeepSeek Coder DeepSeek |
Code-Dense Minimalist | 87% code-trained corpus bias; native Fill-In-The-Middle (FIM) syntax; high instruction adherence with minimal negative constraints. | Degrades rapidly when overloaded with multi-page conversational rules and negative warnings ("do not do X"). |
| OpenAI Codex CLI OpenAI |
Structured ACI Contract | Compact (450 tokens); explicit AGENTS.md scoping hierarchy; clear separation of streaming thoughts vs tool patches. |
Requires strict external sandbox execution; assumes direct human in the loop for ambiguous commands. |
| OpenCode / Interpreter Open Source |
Execution Feedback Loop | Client-server separation; execution error back-propagation; dynamic on-demand Skill loading into sub-agent contexts. | Prone to deep context pollution without strict memory compaction; higher multi-turn token costs. |
| T3 Code pingdotgg / Theo |
Control Plane & Guardrails | Git worktree isolation per task; strict process safety ("3 ways to hurt yourself" rules prevent wild pkill/kill cascades); desktop diff UI. |
Orchestrator rather than an autonomous prompt harness; delegates prompt construction to underlying agents. |
| Exe.dev Fleet Paul Kinlan / Fleet |
Distributed Ledger & Roles | "Beads is the ledger; a message is a photograph"; independent cross-family model review; preemption-aware heavy gate vs fast gate. | Prompt size inflation (roles/common.md was near the 16KB limit); repetition of deterministic rules handled by daemons. |
Claude Code treats the model not as a generic conversationalist, but as an interactive CLI tool operating with user-selected permissions. Key prompt excerpts demonstrate how it enforces safety and clickability:
Key Architectural Takeaway: Rather than dumping 50 edge-case rules into the boot prompt, Claude Code injects <system-reminder> blocks on specific turns (e.g., when Plan mode is activated or a test failure occurs).
OpenAI Codex's system prompt is remarkably brief (less than 500 words), delegating project-specific behavior entirely to the filesystem:
The Fleet prompt architecture solves an existential problem that neither Claude Code nor Codex addresses: autonomous multi-agent coordination without human stalling.
roles/common.md by 65%
Problem: roles/common.md was 15,580 bytes (~3,895 tokens), close to the 16KB deployment limit. Over 40% of its content described mechanisms enforced deterministically by daemons.
Result: Refactored to 5,469 bytes (~1,367 tokens), saving ~2,500 tokens per turn across the fleet while keeping all critical invariants.
file:line)
Inspiration: Claude Code's clickability rule.
Fix: Enforces: "Always reference code locations as path/to/file.ext:line." When reviewers report findings on beads or terminal logs, operators can click directly into the exact line in VS Code or iTerm.
Inspiration: SWE-agent's DISCUSSION → COMMAND pattern.
Fix: Enforced in roles/implementer.md: "State 1 concrete hypothesis → run focused probe → commit code → end turn. Never execute 5+ tool commands without committing progress."
Inspiration: Claude Code's <system-reminder>.
Fix: Instead of having all reviewer and landing rules in every lane's system prompt, lanes.sh spawn injects role criteria only when that specific phase begins.