GitHub Agentic Workflows (gh-aw) v0.90.3 landed October 3. The core: Work Queue plus built-in Ledgers. The Queue keeps todos as a durable backlog, and Ledgers persist state across runs as log, set, map, table, and counter shapes. Task-level model routing, reasoning effort controls, run-wide tool-call budgets, and stronger safe outputs ship together.
The flow changes with it.
Before: Prompt → Agent → done
Now: Queue → Agent → Record → Stop → Resume → Next Work
What arrived: queue and ledger
| Feature | Detail |
|---|---|
| Work Queue | Todos as backlog, Git-backed operator commands (coordinator renamed to work-queue) |
| Built-in Ledgers | 5 shapes (log, set, map, table, counter), cross-run state, compaction via Maintenance jobs |
| Task-level routing | Per-task model picks (AWF task-level model) |
| Budgets | Reasoning effort controls plus run-wide tool-call budgets |
| Safe outputs | Hardened write-request checks (applied in separate scoped-permission jobs) |
GitHub's WorkQueueOps shares the premise. Queues assume interruption, rate limits, and multi-day horizons, with idempotency and retry budgets stated. Multi-day jobs are the design target.
The new normal for long jobs:
- must be stoppable (interruption assumed)
- reruns must stay safe (idempotency)
- retries only inside a budget (retry budget)
Why queues: state and resume beat context
The Durable post said "store before you run." gh-aw brings it as workflow product. Layers 2 (session) and 3 (permanent) from the three-layer memory post arrive named Queue and Ledger.
Giant context windows: seeing a lot this time
Durable state plus resumability: continuing next time
→ the latter bottlenecks long runs
Cost tooling ships along. gh-aw meters inference in AI Credits (1 AIC equals $0.01 USD) and caps runs (max-ai-credits). It reads like cost per success plus execution bundling running inside GitHub Actions. The official guide line is the gem: use deterministic tools when they solve it, agents only when needed.
CodeBridge Mini Lab: attach 4 queue states first
1. Give 1 agent project 4 queue states:
pending, in_progress, completed, failed
2. Fix a retry limit (example: 3 tries, exponential backoff)
3. Add 1 ledger:
- counter: attempt and spend totals
- log: decisions with reasons (auditable later)
- table: per-task status board
4. Run 1 stop test:
- stop mid in_progress, does resume continue
- is rerunning the same job safe (check idempotency)
Per-task model splits from the routing post and approval gates from the security post attach to queue states for completion. The queue is the harness.
Conclusion: the todo list is the agent's body
One line to close.
Prompts start jobs. Queues finish them.
gh-aw v0.90.3 points one way. Not one-shot chats but systems that store state, queue work, and continue across days. One task for today: attach the 4 pending states and a retry limit to your agent. Those 2 start long-running.
Further reading
- The secret of agents that run all night: checkpoints and resume
- Agent memory in 3 layers: context, session, and permanent
- Implementing AI model routing
References
- GitHub gh-aw v0.90.3 Release Notes
- GitHub Agentic Workflows (gh-aw)
- GitHub Agentic Workflows Docs
- GitHub Docs: Cost Management
Go deeper with a course
To design queue, state, and verification loops as a structure, this course builds harness, loop, and graph layers exactly like the Queue, Record, and Resume here.