Frugal Self-Improvement

Agenties can learn from repeated friction, but the improvement loop must be cheaper than the problem it fixes. The current system is passive, local, heuristic-first, and on demand. It does not run a model in the background to analyze itself.

How it works

Step	Mechanism	Cost
1	Append cheap local signals to JSONL.	One file append.
2	Aggregate by signal type and key.	One bounded local file read.
3	Return exactly one improvement proposal.	No model call.
4	Annotate deterministic ROI.	Pure function.
5	User decides whether to create real work.	Human approval.

Signal file

Path

.agenties/self-improvement/signals.jsonl

Signals are append-only records. They are intentionally small so Agenties can track patterns without loading session history or large logs.

Signal type	Example key	Meaning
`repeated_task`	deploy	The same task keeps coming back.
`firewall_block`	batched-work	A spawn firewall block happened repeatedly.
`high_cost_task`	large-review	A task type tends to cost too much.
`permission_repeat`	Bash	The same permission prompt appears often.
`journal_missing`	task-close	A significant task finished without durable context.
`test_failure_pattern`	chatIpc.test	The same test failure pattern repeats.

MCP tools

Tool	Use
`append_improvement_signal`	Append one signal. Pure local write; no model invocation.
`suggest_improvement`	Read aggregated candidates and return one ROI-annotated proposal.

Example

append_improvement_signal({
  type: "firewall_block",
  key: "batched-work",
  metadata: { source: "spawn-firewall" }
})

suggest_improvement({ windowDays: 7, threshold: 3 })

Proposal shape

suggest_improvement returns at most one proposal: the highest-count candidate above threshold.

Response shape

{
  deferred: false,
  proposal: {
    type: "firewall_block",
    key: "batched-work",
    count: 3,
    text: "Recurring firewall block...",
    estimatedCost: "low",
    expectedBenefit: "Fewer spawn blocks...",
    risk: "low",
    validation: "Confirm no firewall blocks for this key in the next 7 days.",
    estimatedTokenSaving: 1500,
    estimatedRoiUsd: null,
    estimatedRoiNote: "USD ROI omitted - token price varies by model and plan."
  }
}

Budget gate

The tool gates on activeRunPressure, not old historical usage. Historical project usage is advisory only for this tool because the analysis is a zero-model local file read.

Warning:Do not block local self-improvement reads because the project has old high cache-read history. Block only when there is real active pressure, such as too many recent spawns or running agents.

Automatic signals

Agenties currently emits firewall_block signals from spawn firewall paths:

Key	Meaning
`prompt-too-large`	The prompt exceeded the spawn firewall size guard.
`batched-work`	The task looked like multiple issues combined.
`usage-high`	The usage firewall blocked a spawn.

← PreviousArtifacts Next →Autopilot