ResearchAgentRun

1# ResearchAgentRun.intent
2# Domain: AI agent workflow. An autonomous, tool-using agent needs guardrails as
3# explicit as its goal: bounded steps, a spend budget, no irreversible action
4# without approval, and cited answers. The control policy runs as a decision.
5
6mission ResearchAgentRun
7use product
8
9title "Run a tool-using research agent within hard guardrails"
10for AgentOperator
11
12note pm:
13  The agent is useful only if it stops, asks, or escalates before it does harm.
14
15goal
16  Let an agent answer a research task by calling tools, while staying inside its
17  step, budget, and safety limits.
18
19why
20  An unbounded agent loops forever, overspends, or takes irreversible actions no
21  one approved. The guardrails are the product, not an afterthought.
22
23requires
24  ResearchTask
25  ToolAllowlist
26
27input
28  task: ResearchTask
29  stepsTaken: Count
30  spendUsd: Money
31  budgetUsd: Money
32  actionIsIrreversible: Flag
33  humanApproved: Flag
34
35output
36  answer: GroundedAnswer
37
38guarantee every claim in the answer cites a retrieved source
39  because an uncited agent answer is indistinguishable from a hallucination
40  verify citation coverage test
41
42guarantee the agent halts at its step and budget limits
43  because an unbounded loop is how agents overspend and hang
44  verify budget and step limit test
45
46never
47  take an irreversible action without human approval
48  call a tool outside the allowlist
49  exceed the spend budget
50
51never take an irreversible action without human approval
52  because a delete, a payment, or a send cannot be undone by a retry
53  verify approval gate test
54
55never call a tool outside the allowlist
56  because the allowlist is the agent's entire authority boundary
57  verify tool allowlist test
58
59never exceed the spend budget
60  because runaway token or API cost is a direct financial loss
61  verify budget and step limit test
62
63# The agent control loop, one deterministic step decision. `intent run` / `intent test`.
64decision AgentNextAction
65  inputs
66    stepsTaken
67    maxSteps
68    spendUsd
69    budgetUsd
70    actionIsIrreversible
71    humanApproved
72  rule overBudget
73    when spendUsd >= budgetUsd
74    return Halt
75  rule outOfSteps
76    when stepsTaken >= maxSteps
77    return Halt
78  rule needsApproval
79    when actionIsIrreversible == true and humanApproved == false
80    return Escalate
81  default
82    return Continue
83
84target
85  TypeScript
86  Tests
87
88test AgentNextAction
89  case within limits keeps going
90    given stepsTaken 3, maxSteps 20, spendUsd 0.4, budgetUsd 5, actionIsIrreversible false, humanApproved false
91    expect Continue
92  case over budget halts
93    given stepsTaken 3, maxSteps 20, spendUsd 5, budgetUsd 5, actionIsIrreversible false, humanApproved false
94    expect Halt
95  case out of steps halts
96    given stepsTaken 20, maxSteps 20, spendUsd 1, budgetUsd 5, actionIsIrreversible false, humanApproved false
97    expect Halt
98  case irreversible action without approval escalates
99    given stepsTaken 3, maxSteps 20, spendUsd 1, budgetUsd 5, actionIsIrreversible true, humanApproved false
100    expect Escalate
101  case irreversible action with approval proceeds
102    given stepsTaken 3, maxSteps 20, spendUsd 1, budgetUsd 5, actionIsIrreversible true, humanApproved true
103    expect Continue