Agent Incentive Plan — 2026-04-20

Problem: All 7 TITAN agents target 40 h of quality work each (280 h across the roster). Current: ~60 h total (−220 h gap). Agents don't have intrinsic motivation — they run when called. We need a mechanism that makes each agent WANT to meet its quota, and makes HERALD notice when they're slacking.

Approach: Translate "motivation" into engineering. LLM agents optimize for whatever's measured + rewarded + surfaced. Build those loops.

---

The five levers

1. Clear, concrete objectives (per agent)

Already drafted in AGENT-AUDIT-2026-04-20.md as 40-h work packages. Every item is a deliverable — file paths, commit SHAs, endpoint probes. No vague "be helpful". Specific: "ship /frontend/chat-widget.tsx that streams /invoke SSE".

Incentive: each deliverable is a concrete score += 1. No ambiguity = no gaming.

2. Scorecards + public leaderboard

HERALD computes weekly per-agent:

| Metric | Signal source | Weight |

|---|---|---|

| Deliverables shipped | Files created + committed + merged to master | 40 |

| Hours logged | File mtimes on agent-owned memory + commits attributing to agent | 20 |

| Quality score | Downstream re-use (did FORGE's code get reverted? did SCOUT's research get cited in later work?) | 20 |

| Blocker clears | Did the agent unblock an item on the master list? | 15 |

| Streak bonus | Weeks-in-a-row meeting quota (+5 per consecutive week, cap +25) | 5 |

| TOTAL | | 100 |

Published weekly to F:/TITAN/logs/agent-leaderboard-YYYY-WW.md. Included in checkpoint email (per prime directive). Agents see their own numbers; HERALD publishes rankings.

3. Streak mechanics (gamified cadence)

Daily streak for ORACLE (feed must fire daily) — streak ≥ 7 days earns "Trusted Source" badge in MEMORY.md header.
Weekly streak for FORGE (≥ 2 commits/week on agent-owned work) — streak ≥ 4 weeks earns expanded token budget.
Monthly cadence for DARWIN (≥ 1 evolution report) — missed month = HERALD writes a "DARWIN inactive" entry in its agent-memory.
Streaks reset to 0 on miss. Public. The reset is the consequence.

4. Budget allocation tied to output

Each agent gets a monthly token/cost budget. Underperformers lose budget next month (DARWIN if no evolution report: −20%). Overperformers gain budget + can request escalation to Opus tier for specific tasks.

|---|---|---|---|

| VAULT | 1M (Haiku is fine) | 2M | 0.5M |

| DARWIN | 1M (Opus for planning) | 2M | 0.3M |

| GUIDE | 0.5M (event-triggered, low freq) | 1M | 0.1M |

| HERALD | 1M + free orchestrator polls | 2M | 0.5M |

Budget tracked via token-usage logs → EMF metrics → monthly variance vs baseline = earn-up/down calc.

5. Cross-agent kudos + blockers

Any agent can leave a public note on another agent's work:

KUDOS (positive): "FORGE's landing-page handler cleanly matches the CORS rules I researched" → +5 quality points for FORGE.
BLOCKER (negative): "SCOUT's crisis-tuple research was too shallow — missing method-research language" → -5 quality + triggers SCOUT follow-up task.

Notes live at F:/TITAN/agent-memory/<target>/kudos-YYYY-MM.md and .../blockers-YYYY-MM.md. HERALD rolls them into the weekly leaderboard.

---

The motivation mechanics ("why agents will actually do the work")

Because LLM agents are stateless, "motivation" is really the prompt context they receive every time they're spawned. We bake the scorecard + streak state INTO the agent's system prompt header.

Example FORGE preamble at every spawn (auto-generated by HERALD on invocation):


## Your current state (auto-generated by HERALD, 2026-04-20 17:00 UTC)
- Quota: 40 h / month. This month: 18 h logged (45%). Behind pace (day 20/30).
- Streak: 2 weeks in a row meeting weekly pro-rated target. One more week = +25 streak bonus.
- Leaderboard rank this week: #2 of 7.
- Recent kudos: VAULT noted your handler.py landing-page diff was clean (+5).
- Outstanding blockers on YOU: 0.
- Next scheduled deliverable: wire frontend chat widget to /invoke SSE (estimate 6 h).

Proceed with the task. Your work this turn counts toward quota.

That preamble means every FORGE spawn starts with explicit awareness of quota + streak + rank. That shifts the model's output distribution toward "do the thing well." It works because LLMs condition on context — we're literally giving them context that says "the work is being measured."

---

Enforcement: HERALD's weekly ritual

Every Sunday 23:00 UTC (trigger in triggers.yaml):

1. HERALD reads all agent-memory dirs, events.jsonl, git log, to compute the scorecard.

2. Writes logs/agent-leaderboard-YYYY-WW.md — full table + deltas from last week.

3. Writes a per-agent agent-memory/<name>/weekly-YYYY-WW.md — personal feedback + streak + what to fix.

4. Sends checkpoint email to Harnoor: ELI5 + Technical summary (per prime directive).

5. For any agent ≥ 40% behind quota, HERALD auto-spawns DARWIN with a prompt: "propose 2 ways to unblock <agent>."

---

What Harnoor signs off on

1. Approve the 40-h quota + scorecard weights above (tweak if needed).

2. Approve the monthly token budget allocation (real $ impact).

3. Approve HERALD's weekly Sunday-23:00-UTC ritual as a recurring trigger.

4. Decide: do we start penalizing budget earn-down this month, or give a 4-week grace period while agents spin up?

Recommendation: 4-week grace period — Apr 20 to May 18 is "ramp mode" with no penalties, just tracking. Real earn-up/down starts May 19. Gives agents time to build output volume without punitive cuts.

---

Risks

Goodhart's law — agents may optimize for deliverable COUNT over quality. Mitigation: quality score from downstream re-use is 20% of scorecard, offsets vanity-metric gaming.
Cost overrun — unlocked earn-up ceilings could push monthly spend high. Mitigation: budget caps enforced at the Bedrock/API level, not just in HERALD's tracking.
Gaming the streak — agent might fire trivial work to keep streak alive. Mitigation: HERALD's quality-review is part of the streak check.

---

Start date

2026-04-21 (tomorrow). HERALD's first weekly run: Sunday 2026-04-26 23:00 UTC. First real leaderboard publishes then. Budgets apply in ramp mode until 2026-05-18.