Gaps And Project Ideas
From learning map to possible builds: run ledgers, failure corpora, tool benches, model routers, promotion ladders, and orchestration labs.
- The fastest way to get better at AI orchestration is to collect failures.
- Backbone builds: a shared run ledger, a failure corpus, a tool fitness bench.
- Workflows graduate up the runtime ladder on pain, not fashion.
Things You Might Be Missing
These are not accusations. They are opportunities worth watching.
1. A Shared Agent Run Ledger
It's common to end up with run metadata in one system. Command telemetry lives in another. Request and step metrics live in a third. There is no shared view across them.
Potential idea:
agent-run-ledger
run_id
system
workflow_or_service
model
tools_called
artifacts_written
trace_path
eval_result
cost
latency
human_interventions
failure_mode
This could become the backbone for learning across all three repos.
2. A Failure Corpus
The fastest way to get better at AI orchestration is to collect failures.
Potential idea:
failures/
task.md
expected.md
actual.md
trace.json
tool_outputs/
diagnosis.md
fixed_by.md
Every bad run becomes training material. It becomes eval material. It becomes design material.
3. Tool Fitness Tests
You have MCP and tool surfaces. The next step is judging whether tools are agent-friendly.
Potential idea:
tool-fitness/<tool-name>/
schema-review.md
eval-cases.json
latency.md
failure-modes.md
prompt-injection-notes.md
- Zendesk MCP
- GitHub
- Linear
- local file and codebase tools
4. Model Routing Scorecards
Not every task needs the strongest model.
Potential idea:
model-router-scorecard.md
task type
cheap model pass rate
strong model pass rate
cost delta
latency delta
recommended default
This is where AI becomes economically scalable.
5. A Workflow-To-Runtime Promotion Ladder
You need a clean rule for when something graduates from prompt/workflow to runner/runtime.
Potential ladder:
Level 1: prose workflow
Level 2: artifact-mediated workflow
Level 3: scripted deterministic runner
Level 4: graph runner with checkpoints
Level 5: scheduled/long-running service
Level 6: self-improving loop with evals and memory
The promotion rule should be pain-based. Only move up when the lower level cannot handle:
- retries
- parallelism
- state
- evals
- latency
Potential Projects
These are ideas worth keeping in a scratchpad.
1. AI Orchestration Lab
A tiny local lab that can run the same task through multiple orchestration patterns:
one-shot
single-agent-with-tools
planner-executor
parallel specialists
generator-verifier
graph runtime
human-in-loop
Output:
runs/<task>/<pattern>/
input.md
trace.json
artifacts/
eval.md
cost-latency.json
notes.md
Why it is useful: it stops you from debating patterns abstractly. You see which one wins on real tasks.
2. Agent Black Box Recorder
A standard trace/artifact bundle for any agent run.
Capture:
prompt or instructions version
model
tool calls
files touched
artifacts written
approvals requested
eval score
final decision
Why it is useful: debugging gets easier. Evals get easier. Trust gets easier.
3. Workflow Builder
A workflow that turns a rough idea into:
workflow.md
manifest.yaml
artifact contract
run metadata contract
eval plan
sync plan
Why it is useful: it makes new workflows less ad hoc. It teaches the system to build itself carefully.
4. PR Review Swarm
Artifact-mediated review:
recon.md
risk-review.md
verification.md
final-report.md
run-meta.json
Why it is useful: it is a clean test for multi-agent value. It does not require a runtime engine first.
5. Tool Fitness Bench
A benchmark for MCP/tool ergonomics.
Test whether the model can:
- discover the right tool
- call it correctly
- use small results
- recover from failures
- avoid unsafe calls
Why it is useful: tool design is a multiplier. Bad tools quietly destroy agent performance.
6. Model Router Playground
Send the same tasks through different models and routes.
- pass rate
- cost
- latency
- tool call quality
- hallucination rate
- the need for human correction
Why it is useful: it builds instinct for when to spend reasoning.
7. Self-Improving Workflow Loop
Not magical self-improvement. A grounded loop:
run workflow
-> collect trace
-> evaluate failure
-> propose prompt/tool/workflow patch
-> human approves
-> rerun eval bank
-> promote if better
Why it is useful: it turns failures into system upgrades. It keeps humans in the promotion path.
8. AI Staff Ops Room
A dashboard-ish concept for agent work would show:
- what is running
- what is blocked
- what needs approval
- what failed
- what got expensive
- what learned something
- what should be retried
Why it is useful: AI systems that become teammates need operations. Chat threads alone are not enough.