GitHub Issues Weren’t Built for AI Agent Orchestration (Here’s What Is)
Developers are duct-taping GitHub Issues, tmux panes, and ad-hoc SQLite databases together to manage Claude Code and Codex agents — but ticketing systems lack the high-frequency event loop, CI gate checks, and structured state that agentic dev work actually demands.
"I've been running an increasing number of local coding agents and I've hit a wall: orchestration and state visibility." That was an engineer on Hacker News in March 2026. His solution was to hack GitHub Issues into a makeshift agent state backend. Nine developers jumped into the comments, each describing their own bespoke workaround. None of them were happy about it.
This is the duct-tape era of AI agent orchestration, and it has a distinct smell.
The most common setup right now is something like: GitHub Issues for task state, tmux panes for visibility, a HEARTBEAT.md file shared between agents for crude coordination, and maybe a hand-rolled SQLite script for gating. This works, barely, when you're running one agent on one repo. Add a second concurrent agent or pick up real commit velocity and the whole thing starts to fray. Issues are sluggish with no native concept of a streaming session or agent heartbeat. Your tmux panes tell you what's happening right now but nothing about why the previous step failed. And the HEARTBEAT.md approach, while clever, is just agents polling a markdown file in a shared directory. It does not scale.
The deeper problem is that these tools were designed for humans reviewing work at human speed. A GitHub issue is a conversation thread. It carries comments, labels, and assignees, but it has no native model for structured handoffs between automated stages. When your agent writes code, runs CI, gets a failure, needs to retry, and eventually submits a PR, there is no single system keeping track of that full loop with structured state at each transition. The agent has to re-read stale context, reason through information it already has, and often repeat work because it has no reliable way to know what already happened.
That is where the token bill comes from. One developer reported spending $900 on AI coding agents in eight weeks. A big chunk of that cost lives in redundant reasoning, in agents re-reading comment threads to reconstruct state they should have been handed directly.
A purpose-built control plane for AI agent orchestration looks different. It models the full dev loop as a sequence of structured stages with explicit gates between them: issue intake with normalised context, code generation, CI execution, feedback consumption, PR submission, review gating, shipping. Each stage produces output that the next stage can consume directly, no reconstruction required. An agent that fails a CI gate gets the failure handed to it as structured data, retries with full context, and moves on. No stale comment threads. No re-reading the whole issue to figure out where it left off.
That is the model behind AgentRail. One structured API for the full dev loop, local-first, source-available. You install it with npm install -g @agentrail-core/cli && agentrail init and it works alongside Claude Code, Codex, Cursor, or whatever agent you're running. It does not try to replace your agents. It gives them the structured scaffolding that generic ticketing tools were never designed to provide.
The token numbers are measurable. Against plain Codex on the same tasks, AgentRail shows 47% fewer total tokens consumed and 93% fewer reasoning tokens. The reasoning reduction is the more telling figure. It reflects how much agent effort was previously going into reconstructing state rather than doing actual work.
There are real limits worth naming. AgentRail is early. The source-available model means you can read the code and adapt it, but the ecosystem around it is still thin. If your workflow is genuinely simple, one agent, one repo, low velocity, the duct-tape approach may be fine for now. The overhead of a structured control plane earns its keep when you're running multiple agents concurrently or when you care about token costs at scale.
If you're past that threshold, or approaching it, take a look at https://agentrail.app. The gap between what developers are building by hand and what a control plane should provide is real, and it is already costing people time and money.