Capture the complete "train of thought" of agent work as first-class artifacts.
When an agent completes a task today, the only artifacts are code changes, commit messages, and PR descriptions. The rich context of how the work happened disappears: why approach A was chosen over B, what dead ends were explored, what assumptions were made.
Agent Trajectories captures this missing context as structured, searchable, portable records that travel with the code.
A trajectory is the complete story of agent work on a task:
- Chapters - Logical segments of work (exploration, implementation, testing)
- Events - Prompts, tool calls, decisions, messages between agents
- Retrospective - Agent reflection on what was accomplished, challenges faced, and lessons learned
- Artifacts - Links to commits, files changed, and external task references
Works with any task system: Beads, Linear, Jira, GitHub Issues, or standalone. Trajectories are a universal format—like Markdown for documentation.
- File system (default) -
.trajectories/directory, git-friendly - SQLite - Local indexing and search
- PostgreSQL/S3 - For teams and archival
- Markdown - Notion-style pages for documentation
- Timeline - Linear-style chronological view
- JSON - Full structured data for tooling
- Complements claude-mem for observation-level memory
- Integrates with agent-relay for multi-agent messaging
Instead of guessing at intent from 500 changed lines, reviewers can:
- Read the trajectory summary
- See what alternatives were considered and rejected
- Understand the agent's confidence level
When a bug surfaces months later:
- Query the trajectory for the commit that introduced the code
- See original requirements and edge cases considered
- Understand the context that led to this implementation
Over time, trajectories become a searchable knowledge base:
- "How have we solved caching problems before?"
- "What libraries did we evaluate for X?"
- "Why did we choose this architecture?"
# Start tracking a task
trail start "Implement auth module"
# View current status
trail status
# Record a decision (reasoning optional for minor decisions)
trail decision "Chose JWT over sessions" \
--reasoning "Stateless scaling requirements"
# Complete with retrospective
trail complete --summary "Added JWT auth" --confidence 0.85
# List all trajectories (with optional search)
trail list
trail list --search "auth"
# Export for documentation (markdown, json, timeline, or html)
trail export traj_abc123 --format markdown
trail export --format html --open # Opens in browserTrajectory = the complete path an agent takes through a task Trail = what's left behind for others to follow
You don't see the whole trajectory in real-time, but you can always follow the trail.
The CLI is called trail because that's what you're doing—leaving a trail of breadcrumbs through your work. Future agents and humans can follow this trail to understand not just what was built, but why it was built that way.
Both agents and humans—but differently.
Agents use trail commands to record their work as they go:
# Agent starts work on a task
trail start "Add rate limiting to API"
# Agent records key decisions as it works
trail decision "Token bucket algorithm" \
--reasoning "Better burst handling than fixed window"
# Agent completes with reflection
trail complete --summary "Added rate limiting" --confidence 0.9This can be invoked programmatically by AI coding tools, or agents can learn to call trail as part of their workflow.
Humans use trail commands to understand and review agent work:
# List and search past work
trail list --search "authentication"
# See trajectory details and decisions
trail show traj_abc123 --decisions
# View in browser
trail export traj_abc123 --format html --open
# Export for code review
trail export traj_abc123 --format markdownThe trail bridges the gap between agent work and human understanding:
Agent works → Records decisions → Completes trajectory
↓
Human reviews → Follows the trail → Understands the "why"
Without the trail, humans see only the code. With it, they see the reasoning.
Trajectories power a broader vision: a knowledge workspace for agents—like Notion, but for AI.
┌─────────────────────────────────────────────────────────────────┐
│ AGENT WORKSPACE │
├─────────────────────────────────────────────────────────────────┤
│ 📚 Knowledge Base 🛤️ Trajectories │
│ ├── Architecture docs ├── Active work │
│ ├── Code patterns ├── Recent history │
│ └── Conventions └── Searchable archive │
│ │
│ 🧠 Decision Log 📋 Pattern Library │
│ └── Why things are └── How to do things │
└─────────────────────────────────────────────────────────────────┘
When an agent starts a new task, it can query the workspace for:
- Relevant past trajectories
- Applicable patterns and conventions
- Related decisions
- Potential gotchas from retrospectives
┌─────────────────────────────────────────────────────────────────┐
│ AGENT-TRAJECTORIES (Layer 3) │
│ Task narratives, decisions, retrospectives │
│ ▲ │
│ │ aggregates │
│ CLAUDE-MEM (Layer 2) │
│ Tool observations, semantic concepts │
│ ▲ │
│ │ captures │
│ AGENT-RELAY (Layer 1) │
│ Real-time messaging, message persistence │
└─────────────────────────────────────────────────────────────────┘
Each layer is independent and can be used alone, but together they form a complete agent memory stack.
{
"id": "traj_abc123",
"task": {
"title": "Implement user authentication",
"source": { "system": "linear", "id": "ENG-456" }
},
"status": "completed",
"chapters": [...],
"retrospective": {
"summary": "Implemented JWT-based auth with refresh tokens",
"decisions": [...],
"confidence": 0.85
}
}Trajectories are stored as .trajectory.json files (machine-readable) with auto-generated .trajectory.md summaries (human-readable).
"The trajectory is as valuable as the code."
As AI agents write more code faster than ever before, a critical gap emerges: we're shipping code without understanding. Trajectories close this gap.
Without trajectories, agent-generated code becomes a black box:
| Problem | Impact | How Trajectories Help |
|---|---|---|
| Silent assumptions | Bugs hide in undocumented edge cases | Decisions and reasoning are captured explicitly |
| Inconsistent patterns | Each agent reinvents approaches | Past solutions are queryable, patterns emerge |
| Lost context | Nobody knows why code exists | The "why" lives alongside the "what" |
| Review theater | PRs approved without real understanding | Reviewers see the full decision history |
| Debugging blind | Hours spent reverse-engineering intent | Original context is one query away |
Trajectories create a virtuous cycle that compounds over time:
More trajectories → More extracted knowledge → Better agent context →
Better decisions → Better retrospectives → Richer trajectories → ...
Each completed task makes future tasks easier:
- Agents make fewer mistakes by learning from past gotchas
- Decisions are more consistent across the codebase
- Onboarding new agents (or humans) becomes instant
- Institutional memory persists even as team members change
As agent usage scales, trajectories become essential infrastructure:
Today (1-2 agents):
- Nice to have for code review
- Helpful for debugging
Tomorrow (5-10 agents working in parallel):
- Critical for coordination
- Required for understanding who did what and why
- Enables agents to learn from each other
Long-term (agents as primary contributors):
- The authoritative record of how the system evolved
- Training data for project-specific agent improvements
- Audit trail for compliance and security review
Agent-generated code faces a trust problem. Developers hesitate to ship code they don't understand. Trajectories solve this by making agent reasoning transparent:
- Confidence scores tell you when to scrutinize more carefully
- Decision logs show trade-offs were considered
- Retrospectives surface known limitations and risks
- Challenge documentation reveals what was hard (and might break)
The result: teams can ship agent code with the same confidence as human-written code—because they understand it just as well.
This project is in early development. See PROPOSAL-trajectories.md for the full design document.
MIT