AIM is a way to run agentic software delivery without losing the thread. It gives AI work a clear loop instead of a prompt spiral.
Without a method, agentic development usually breaks in predictable ways:
AIM fixes that with clear roles, gates and ownership.
Roles are played in sequence for each Done Increment. The loop repeats until the Epic is complete.
Owns the Epic. Defines value, scope, non-goals and acceptance criteria. Decides when the Epic is complete.
Translates the Epic into one concrete Done Increment. Limits scope. Validates the increment after review.
Implements exactly one Done Increment. Focus is on end-to-end value, not partial work.
Reviews for correctness, edge cases and technical risk. Provides concrete, actionable feedback.
Approvals only matter at three points. Everything else is signal, not permission.
PO frames the Epic. Scope and value boundary agreed.
TDO proposes the next single Done Increment. PO approves or adjusts scope.
Dev and Reviewer report progress and risk. No approval needed unless escalation occurs.
TDO presents the completed increment. PO accepts or requests adjustment.
Choose one path:
/aim start "EPIC: <desired user outcome>"
Mode: Strict
Cost profile: Cost ControlUse Cost Control for ordinary low-risk work. Use Deep when the work touches trust, data correctness, deployment, migration, security, or public APIs.
Pauses at the meaningful hard gates. Default for new teams or high-trust-sensitive work.
Continues through increments automatically unless an escalation condition is hit. Same gate logic, faster throughput.
AIM works with the tools you already use. The method stays the same across all of them.
Use the packaged aim agent in .github/agents/. Add .github/prompts/ for Copilot-style command helpers.
Use CLAUDE.md plus the shipped starter files in .claude/commands/ and .claude/agents/.
The repo is the AIM contract. The Codex skill adds the /aim launcher plus bootstrap help.
Important: .github/agents/aim*.agent.md are part of the AIM instruction layer, not just Copilot decoration. Install them regardless of which adapter you use.
Cost profile controls runtime depth, not approval flow. Gates and escalation rules are the same regardless of which profile you choose.
Normal AIM with progressive context loading and compact gates. The default for most work.
For low-risk, reversible cleanup or narrow documentation changes. Escalates to Standard or Deep when risk appears.
For trust-sensitive, data correctness, deployment, migration, security, or public API work. Broader inspection, stronger review.
Cost Control is not weaker AIM. It is AIM with a smaller runtime budget and a clear rule to escalate when risk appears.
Add AGENTS.md, docs/workflow/agile-iteration-method.md, and .github/agents/aim*.agent.md to the target repo.
/.aim to .gitignoreThe .aim/ folder is runtime state. AIM creates it automatically on first start.
AGENTS.md repo-awareAdd your stack, verification strategy, and deployment constraints. Vague rules produce vague behavior.
EPIC: <desired user outcome>
Mode: Strict
Cost profile: Cost ControlFor the full installation guide: docs/workflow/install-aim-1.6.md