Coding Agent
16 min readImplement approved user stories in your linked Git repo — configure runs, batch Implement All, context, tokens, and Insights.
Coding Agent
Coding Agent implements your approved user stories directly in your linked Git repository — without you leaving Forge. It clones your repo, reads your full journey context (architecture, PRD, acceptance criteria), writes the code, and pushes a branch with a pull request ready to review.
You pick the story. Forge does the implementation.
Prefer interactive control? Use with the same stories and Suggested AI Models.
What You Get
| Outcome | What you get |
|---|---|
| Code implemented | Against your linked Git repo |
| Dedicated branch | forge/wo-{id} — pushed automatically |
| Pull request | Opened when your provider allows |
| Live logs | Streaming as the agent works |
| Full traceability | Runs tied to the story with reports and logs |
Before You Start
Three things must be in place before Coding Agent will run:
1. Application Context saved
Your project must have stories under Application Context → Stories.
2. Repository linked
Link a repo under the ALM Repos tab using a Git connector with the full repository name. A URL-only link is not enough — the connector must have push and PR create permissions.
3. Story is coding-eligible
Stories marked as manual-only do not show the Coding Agent tab.
Run Coding Agent on a Story
Step 1 — Open the story
Go to Application Context → Stories → open the story you want to implement.
Step 2 — Open the Coding Agent tab
Click the Coding Agent tab on the story, or use the card menu → Run Coding Agent.
Step 3 — Configure the run
| Field | What it does |
|---|---|
| Repository | The linked repo Forge will clone and push to |
| Instructions for AI | One-off guidance for this run only (~2000 chars) |
| Commit Prefix (Advanced) | e.g. JIRA-123 prepended to commit messages |
| Skip testing (Advanced) | Skips the test phase — only use if CI will still gate the PR |
| Skip code review (Advanced) | Skips the AI review phase |
| Base Branch (Advanced) | The branch Forge forks from |
Review the token estimate shown — this is a planning guide, not a bill.
Step 4 — Click Implement with Coding Agent
The agent starts. You'll see:
- Status: Coding Agent Running
- Live Logs streaming in real time — file reads, commands, AI reasoning as it works
- Cancel Execution button if you need to stop the run
Closing the tab does not stop the run. Reopen it to reconnect to an active session.
Step 5 — Review the result
When finished, the result panel shows one of: PR Created, Changes Pushed, or Implementation Complete.
From here you can:
| Action | Where |
|---|---|
| Open the pull request | View Pull Request |
| See phases, AC coverage, files changed | Result panel |
| Download evidence | Download Report / Download Logs |
| Run again | Run Again — prior runs are listed in the configure view |
The branch in GitHub will show as forge/wo-{id} with the commit message [WO-XXX] Story title — pushed by the Forge Coding Agent.
What Context the Agent Uses
The agent does not start from a blank prompt. Each run assembles:
| Source | How it is used |
|---|---|
| User story | Title, description, acceptance criteria, priority, points |
| Architecture | Approved architecture options from the journey |
| PRD / BRD | Product and business requirements context |
| UI Design | Page and layout summary when UI Design was approved |
| Repository | Live clone of the linked repo at the base branch you select |
| Coding Agent Instructions | Persisted per-story guidance set on the story Details tab |
| Instructions for AI | One-off extras added at launch for this run only |
| Prior feedback | Previous run or council feedback when available |
Persist Instructions vs One-Off Instructions
Coding Agent Instructions (story Details tab) — saved permanently on the story. Applied on every run including retries and batch. Use this for scope rules, file paths to avoid, or coding standards that should always apply.
Instructions for AI (launcher) — applies to this run only. Use for temporary overrides or one-time context.
Run on Multiple Stories — Implement All
When several stories are ready, use batch mode instead of running them one by one.
Step 1 — Open batch mode
Stories toolbar → Implement All → Batch Coding Agent
Step 2 — Select stories
Use Select All ({n} actionable) or pick individually. Batch automatically skips stories that are completed, removed, in review, blocked, or marked manual-only. If a selected story has unresolved dependencies, batch can auto-include the blocking stories.
Step 3 — Choose a mode
| Mode | Use when |
|---|---|
| Sequential (respects dependencies) | Stories block each other — safer default |
| Parallel (concurrent) | Stories are fully independent — faster |
Step 4 — Review and launch
Review the story list, estimated duration, and token estimate. Click Implement with Coding Agent ({n} Stories).
You can minimize the batch panel — it continues in the background. Go to Insights → Coding Agent to watch live logs for any story in the batch.
If the batch fails partway through, use Resume Remaining to continue from where it stopped, or Start Fresh to rerun everything. Download the batch report at any time.
Token Estimates
Shown on the story Details tab, the single-story launcher, and the Implement All panel.
Estimates are calculated from prior coding-agent analytics scaled by story complexity — low (0.6×), medium (1.0×), high (1.5×) — or a platform baseline when no history exists. Actual usage varies with repo size, number of retries, and whether testing and review phases run.
Platform admins can see realized token counts and cost in Insights and run results.
Insights
Track Coding Agent activity per project under Application Context → Insights → Coding Agent.
| Metric | What it means |
|---|---|
| Total Runs | All executions — succeeded and failed |
| Success Rate | Percentage of successful runs; stories completed by the agent |
| Avg Duration | Mean runtime and average files changed per run |
| Remaining Estimate | Projected work for stories not yet run |
The Recent Executions list shows individual traces, diffs, reports, and a View Live Logs link for each run.
Empty: No Coding Agent Runs Yet. Platform admins: Admin → Observability → Coding Agent for cross-tenant views.
Suggested AI Models
On story Details, expand Suggested AI Models for Recommended and Budget Alternatives by complexity (low / medium / high).
For IDE / MCP only. The Forge Coding Agent launcher does not use this picker — the worker uses platform defaults.
MCP get_work_order / get_next_work_order include the same recommended_models. See .
Forge vs IDE
| Coding Agent (Forge) | Open in IDE / MCP | |
|---|---|---|
| Best for | Hands-off and batch runs | Interactive steering mid-implementation |
| Where | Story Coding Agent tab, Run Coding Agent, Implement All | Open in IDE; MCP in Cursor / VS Code / etc. |
| Who writes code | Forge worker | Your local agent |
| Model | Platform defaults (no launcher picker) | You choose; Forge suggests via Suggested AI Models |
| Output | forge/wo-{id}, push, optional PR, report/logs | Commits / PRs from your IDE workflow |
Troubleshooting
| Symptom | What to try |
|---|---|
| No linked repositories found | Repos → link with connector + full name; Test the connector |
| Coding Agent tab missing | Story may be manual-only (requires manual implementation) |
| AI timeout / AI error | Simplify Instructions for AI; Run Again |
| Test failure | Fix tests in a follow-up, or Skip testing if CI will still gate the PR |
| Auth / push / PR failed | Check connector permissions (push + PR create) and base branch |
| Batch stuck / partial | Resume Remaining or Start Fresh; check Insights live logs |
| Blank / stalled Progress | Reopen the tab to reconnect; confirm workers are healthy |
| Unexpected files changed | Tighten Coding Agent Instructions (scope / paths to avoid) |
Expand Why did it fail? on the result panel for category-specific hints.
Best practices
| Practice | Why |
|---|---|
| Persist rules in Coding Agent Instructions; use launcher text for one-offs | Consistency across retries and batch |
| Prefer Sequential until dependencies are clean | Avoid blocked or out-of-order PRs |
| Link the repo your team already reviews | PRs land where reviewers already work |
| Start with smaller stories, then larger ones | Faster feedback on agent quality |
| Skip testing only when CI still validates | Don't ship untested changes by accident |
| Use Open in IDE for messy, exploratory stories | Human-in-the-loop when the path is unclear |
| Approve architecture (and UI Design when relevant) first | Better context → better PRs |
Related
- — Stories, Repos, Insights
- — tenant-wide analytics (not per-project Coding Agent KPIs)
- — IDE workflow +
recommended_models - — capabilities, Figma, chat, upload, and
.figdownload - — GitHub / GitLab / Bitbucket